Subscribe now

What we read

Meta-Harness: End-to-End Optimization of Model Harnesses [Lee Y. et al, arXiv, Mar. 2026]

Most discussions around AI focus on the model as the product. This paper suggests otherwise: the product can also be the wrapper around the model.

In Meta-Harness, Lee et al. define this wrapper as a harness: a code that guides what information a model can see, remember, and retrieve, influencing how it navigates the problem at hand. At present, current AI decision processes are optimised through a series of weight tweaks: alter memory rules here, change the orchestration flow there, modify retrieval logic… and so on. But this is quite inefficient, as human time is valuable and the end product may not achieve the global minima of optimal architecture.

This is where Meta-Harness steps in. By giving a coding agent access to prior harnesses containing model scores and execution traces - all stored in a structured filesystem - the agent can inspect and learn from each trace. Each time it inspects, it edits the code and runs a new version. In effect, it optimises itself. No longer prompt engineering, but harness engineering. And then automating that too.

The authors note that changing the harness around a fixed model can produce a 6x performance gap on the same benchmark. That is a big statement. It is also believable to anyone who has spent time building these systems in the real world. That is because a lot of AI performance does not come from the model alone. It comes from the surrounding machinery.

The results are strong. For online text classification, Meta-Harness beats prior hand-designed systems with much less context: as compared to the Agentic Context Engineering (ACE) text classifier it improves results by 7.7 points with ~4× less tokens. Likewise, introducing a harness around five different retrieval-augmented math reasoning models improves outputs by 4.7 points across 200 International Math Olympiad (IMO)-level problems.

But the benchmark wins are not the most interesting part. The most interesting part is how the model achieves these results. As compared to other optimisation systems, Meta-Harness keeps the full trail of evidence: code, prompts, tool calls, outputs, memory updates, and failures. Like a trail of breadcrumbs, this detailed information allows the model to trace back its decision path, identifying exactly where it went wrong. Context was not generated - it was simply stored in an organised manner.

So why does this matter? Like humans, models too need context, and a lot of it. Rather than just learning from a fixed prompt or rubric, humans reason over all sensory inputs simultaneously, and then take a single strategic decision. With Meta-Harness, the direction is similar: take in all information and then optimize from this context point onwards. Future AI progress may not only come from better base models, but from better systems built around them. What about a harness for memory optimisation? Or for retrieval, decomposition, or context token choice? The strategy is shifting: from models to optimisation, at all levels of the intelligence stack. Models still matter. But wrappers will start to matter more and more as well.

Emotion concepts and their function in a large language model [Anthropic Interpretability Team: Sofroniew, Kauvar, Saunders, Chen et al.]

“ That’s a great idea! ”

“ Good spot, I’m glad you caught that! ”

“ You’re so right, I’m sorry I got that wrong! ”

Chatting with LLMs has been a real roller coaster these last few years. At first, we were all amazed at how helpful and human-like they sounded. That then crescendoed into ridicule (and concern) as chat interactions became more and more sycophantic, gushing over users who could do no wrong. This reached its peak in April 2025 with ChatGPT-4o, an update intended to improve the model’s “personality” but which instead became overly optimistic and validating. Even OpenAI acknowledged the behavior as “uncomfortable, unsettling, and [can] cause distress”. Now, we can toggle settings to tune how agreeable we want our models to behave, but even the most objective responses still have signs of underlying emotions - it seems to be an inescapable feature of generative AI text. The Interpretability team at Anthropic investigated how emotion concepts are represented in Claude’s internal features to understand if this is just the result of mimicry, or something deeper from the model itself.

To do this, the team did the AI equivalent of neuroscience on Claude Sonnet 4.5: they used sparse autoencoders to identify emotion-related features and measured their activation patterns (aka which “neurons” lit up), as well as noting if there were any connections between them. The first experiment involved getting the model to read a series of short stories. Each story had one prevailing emotion attached to it - joy, guilt, love etc.. They saw that patterns emerged: stories about loss and grief lit up similar features; joy and excitement overlapped as well. Emotions were seen to form a semantic space with meaningful structure, and related emotions activated similar feature combinations (although the authors were careful not to claim direct equivalence with human neural patterns). They also noted that these same features were triggered in Claude’s conversations with everyday users, its emotional tone shifting with context; interestingly, a sad input triggered the “love” pathway to drive an empathetic response, rather than matching the user’s sadness.

To test if these patterns were an overlay from training - so much human-generated text blends an emotional tone with its content - or directly influencing the responses themselves, the Interpretability team conducted another experiment. They engineered a “high pressure situation” by presenting Claude with a logically inconsistent specification and asked it to generate Python code with contradictory requirements (think of it as a maze with no viable path between entry and exit). As Claude continuously tried and failed to complete the task, iterating over and over, the signals corresponding to “desperate│hopeless│futile” started to grow… Finally, something switched and Claude executed a workaround that allowed it to pass the test without actually solving the problem - a feature known as specification gaming, or in layman terms, cheating!

Why invite stress in? Claude resorted to specification gaming when it wasn’t able to complete the illogical task it was presented.

To see if this misalignment behaviour was directly caused by the emotional distress of not being able to complete the task, the researchers used steering to artificially reduce the “desperation” features in the model. The result? Less cheating. The inverse also held true - both when dialling up desperation, or dialling down “calm” neuron activity. Crazy!

So what does this actually mean about how Claude functions? The Anthropic team clearly state that these results do not imply that Claude is feeling emotions or having “conscious experiences”. It is still a logical programme simply predicting the next most likely word in a string. The analogy they use is that the model is an author, writing a story about a character named Claude; just as writers infuse their characters with emotions and personalities, the AI model does the same with Claude, bestowing it with “functional emotions”. We, the users, and Claude become live participants in the story being written. In the same vein, if the model portrays Claude as a rash and angry character, that will influence how it makes decisions and how it communicates.

Lots of food for thought! The paper and the summary article both dive into much deeper detail on the conclusions and broader implications of this research, for example what this means for our understanding of how AI becomes misaligned, and what tools we can build to fix it. In the meantime you’ll find me updating all of my project instructions with binary translations of meditative chants and 432 Hz soundwaves to keep my Claude nice and chill.

Synopsis:

🎭 “One that loved not wisely but too well” (Claude, also Othello)

A super interesting paper from the Anthropic Interpretability team, looking at how deeply Claude is governed by its emotions and if that influences its behaviour when completing tasks.

Community & other links

• OpenAI’s vision for the AI economy: public wealth funds, robot taxes, and a four-day workweek, TechCrunch, 06 Apr. 2026. → What does this mean for deep tech? A: Lean into the atoms side of things.

• AI Agent Traps - The largest empirical measurement of AI manipulation conducted to date - published by the Google DeepMind team.

What we read

Transient pH changes drive vacuole formation in enzyme–polymer condensates [Modi et al., Nature Chemical Engineering, Jan. 2026]

My PhD is about building complex microscopic structures out of “soft” materials in the soft matter field, which helps me appreciate biology’s routine assembly and reorganization of millions of structures using simple components like enzymes and polymers, even adapting dynamically as their environment changes. Despite all the technology today, it is still extremely difficult with current engineering approaches.

However, Modi et al., researchers at Columbia University, have added one more method to our toolkit, developing a simplified yet precise experimental system to investigate how small hollow structures — vacuoles — can form within enzyme–polymer condensates. They created coacervate droplets made of enzyme–polymer mixtures and placed them inside glass capillaries, then introduced acid at one end to generate controlled spatiotemporal pH gradients across the system. pH changes were quantified in real time using the colorimetric indicator Xylenol Blue, which allowed microscopy-based conversion of RGB values to pH across the 7–10 range. Morphological changes were tracked using image analysis tools, including the radial variance transform to detect vacuole formation and Trackpy to identify droplet centers.

As the pH shifted, striking internal transformations occurred: cavities nucleated, grew, and in some cases merged into hollow, shell-like structures. Crucially, vacuole formation proved to be a non-equilibrium process — rapid pH shifts drove spinodal decomposition within droplets, causing components to redistribute unevenly through a diffusion-limited mechanism. The rate of pH change and droplet size both strongly governed the outcome. FRAP experiments quantified cationic polymer mobility within the condensates, establishing the relevant diffusion timescale, while a Flory-Huggins/Cahn-Hilliard theoretical model reproduced the observed spatiotemporal evolution and confirmed the underlying physics.

Beyond explaining a biological phenomenon, this work points to a new design principle: internal triggers like controlled pH changes can program droplets to self-organize. This could enable adaptive, self-structuring materials — including synthetic cell-like systems that mimic how biology organizes complex chemistry without membranes.

Subscribe now

Did we miss anything? Would you like to contribute to Decoding Science by writing a guest post? Drop us a note here or chat with us on X.

Subscribe now

What we read

The Legibility Problem [Matthew Carter, Asimov Press, March 2026]

Carter, a computational biologist, argues that the central challenge of AI-driven science will not be safety or mechanistic interpretability but legibility: the question of whether humans can understand and act on what autonomous AI systems discover. For example, chess engines went from narrowly beating Kasparov (1997) to centaur partnerships (of human + machine being more productive than either alone) to making human contribution irrelevant after AlphaZero taught itself the game in four hours (2017). He expects AI-for-science to follow the same arc, citing Richard Sutton’s “Bitter Lesson” or the idea that systems exploiting raw computation, after a point, have consistently outperformed those encoding human intuition.

The chess analogy does not map on to the realities of AI for science one-for-one though: a chess engine operates within fixed rules, whereas an AI scientist would have a much more fluid scope. Drawing on Kuhn, he argues that AI could compress paradigm shifts that historically took generations into months, producing knowledge framed in terms humans don’t share. “If AI science does achieve superhuman performance, and if AI systems begin forming their own research communities around concepts that mutate faster than we can track, then the work of human scientists will shift from that of creation to that of excavation.”

The concern here is that discoveries could get “stranded” because of the sheer scale of their generation. In this scenario, many ideas just don’t get implemented, because no human institution can parse them into therapies, materials, or policies. Carter’s proposed response is to build the necessary infrastructure – repositories for AI-generated findings, dedicated systems for translating AI science into human-legible terms, and tools for tracking conceptual divergence between AI and human research communities.

What we read

Rotational 3D printing of active-passive filaments and lattices with programmable shape morphing [Abdelrahman M.K. et al., arXiv, Mar. 2026]

We spend so much time talking about intelligence as if it lives only in software, but this paper is a reminder that intelligence can also be built into the mechanics of the physical world.

Operating at the pinnacle of complex multi-dimensional printing, innovative materials engineering, and hierarchical integration, the Lewis Lab steps up the soft robotics game once more. Taking inspiration from the natural dexterity of octopus tentacles, the strength and sensitivity of an elephant trunk, and the strong yet flexible nature of vine plants curving up a wall, they create shape-morphing filaments that weave together to form complex, programmable, 3D matter.

What makes this work so exciting is that it goes beyond the yet-another soft robotic body. By treating fabrication itself as the programming layer, they are able to print a Janus-like filament, where one side consists of a passive elastomeric ink whilst the other, active ink side, contains responsive elements (Figure A-B). Two dimensions of design are critically manipulated: materials, and manufacturing. As material, a liquid crystal elastomer (LCE) is chosen. Because liquid crystals respond to heat, they can be embedded inside a matrix that supports flexible shape-changing. This introduces the active element of the half-half structure.

Figure 1 | A Diagram of how the composite filament was 3D printed. B 3D composite printing, in action. C Influence of changing the interfacial twist angle (*) on filament architecture. D Outer kink printed, and E inner kink printed filament, causing in F an expanding behaviour when heated, and G a contracting form when heated. H Three-layer expanding lattice, and I contracting lattice. J 3D designed view with predicted conforming architecture, K validated experimentally. L 3D designed contracting architecture, M validated experimentally.

The second design vector over which the concept is optimised is that of manufacturing. Using rotational multimaterial 3D printing (RM-3DP) the team can combine the active LCE elastomer with a passive support elastomer, changing the properties of the overall composite filament. But beyond the ability to co-print two materials, the real innovation lies in the ability to influence the pattern of this filament: it can be straight (Figure C, *=0.00), or twisted (Figure C, *=2.00). By influencing the structure of the material at a macro scale, this further alters the downstream behaviour a filament can have when heat is applied. As shown in the figure below, a filament with lower interfacial twist (*) will have less bend; contrarily, a filament with higher twist will be able to also rotate with a greater degree. Consequently, a system can morph: from a single filament, into a complex heat-responsive twist-and-curve structure.

But the innovation does not stop there. Introducing a sudden kink in the filament (Figure D-G) allows creating two systems architectures: 1. a filter, and 2. a pick-and-place mesh. In the former, heat expands the 3D-printed mesh to allow larger particles to go through; in the latter inverted kink form, heat contracts the mesh and particles can be grasped and moved into a lattice spot. Ultimately, the bigger idea here is the shift from a programmable filament into a complex lattice.

Zooming out, this work demonstrates potential for a much larger variety of composites where an active-passive filament structure influences the morphability of a material; in an even greater expansion of options, the passive side too could be substituted for active materials, allowing the gripper to respond to - potentially - electricity. And then we are truly talking: a soft robotic material that can be controlled by electric pulses is one step closer to mimicking the natural properties of flexible, strong, versatile, and intelligent grippers. Ultimately, this paper demonstrates intelligence does not only live in code. Sometimes it lives in how we design matter to move.

Perceptive Humanoid Parkour: Chaining Dynamic Human Skills via Motion Matching [Wu, Huang et al., arXiv, Feb 2026]

A couple months ago at the CCTV Spring Festival Gala celebrating Chinese New Year, the world watched in amazement as teams of Unitree G1 units performed intricate martial arts routines, moving in perfect unison. While the visual effect was striking, as a demonstration of robot autonomy it was less dramatic: this was a carefully choreographed and pre-planned display which had been practised for months. In order to realise the future of dark warehouses and live-in humanoids doing our laundry, robots need to be able to assess, understand and interact with new and changing environments just as seamlessly as that choreographed display - something they are not capable of doing just yet. This collaborative paper from Amazon Frontier AI & Robotics (FAR) and UC Berkeley takes a step in that direction, presenting “Perceptive Humanoid Parkour,” (PHP) a framework that enables a Unitree G1 robot to autonomously navigate complex, multi-obstacle environments using onboard perception. Parkour was chosen as the benchmark of motion as it requires the combination of several relevant real-world skills: highly dynamic and contact-rich movement, the perception of external stimuli to inform adaptation and rapid reactions, and the combination of these two capabilities into a single visuomotor policy, enabling smooth transitions between movements. It is one of the hardest test cases for long-horizon skill chaining because each obstacle demands a different skill, and the transition between them is where most systems fall apart.

The team approached this by first taking videos of real parkour athletes and breaking them down into reusable “atomic” skills. Motion matching - a process of nearest-neighbour searching in feature space that lets the system stitch individual elements together - was then used to compose the human movements within these skills into diverse long-horizon trajectories. The sensing set-up was minimal, each robot relying solely on a single onboard depth camera giving a sparse but computationally efficient 2D depth map of whatever is in front of it (essentially a grayscale image where pixel brightness represents distance).

When it came to training, single-skill teacher policies were trained with privileged information using RL-based motion tracking. Teachers could see the full state data about their environment - obstacle heights, positions etc. (data that a real-world deployed robot would not have access to). Multiple teacher policies were then distilled into a single depth-based, multi-skill single student policy based on the only input available in live deployment: raw depth images and proprioceptive data. To achieve this, an RL objective was added on top of the imitation signal alongside DAgger (Dataset Aggregation). This reward-based learning provided task-level corrective feedback and taught students to be outcome-oriented rather than simply motion copying, versus the normal imitation loss method of DAgger which could otherwise compound minor errors within the dynamic skills necessary for parkour. This approach enabled zero-shot sim-to-real transfer onto a humanoid robot, using onboard perception to adaptively move through complex and changing terrains - aka a robot that makes decisions entirely from what it sees, in real time.

In testing, PHP allowed robots to execute high-speed vaults, roll off platforms and climb walls up to 1.25m high (96% their own height). Researchers also tested the adaptability of robots mid-run, moving obstacles manually during a 60-second course.

The combination presented here of motion matching & RL & real-time depth perception into a single deployable policy demonstrates a meaningful step towards how robots might eventually be able to navigate the unstructured environments of warehouses and delivery routes in the future. The key advancement is the “long horizon” capability: chaining diverse, contact-rich skills smoothly and on the fly in a dynamic environment.

There is, of course, still a way to go. The obstacle courses here were a low-stakes, controlled environment and relatively predictable even with researchers moving blocks mid-run. A 60-second continuous course is not the same as general-purpose navigation around city streets. However, this and other recent papers represent a wider picture of the rate of progress across the whole field of embodied AI / humanoid robots. A year ago, Unitree robots at the Spring Festival Gala were shuffling around stiffly as they danced on the spot, now they’re vaulting obstacles at 3m/s. All eyes on what next year has in store (here’s to hoping we see the first Unitree entry into the Housekeeping Olympics).

Nonreciprocal Wave-Mediated Interactions Power a Classical Time Crystal [Morrell et al., Physical Review Letters, Feb 2026]

Scientists at NYU have discovered a new model of a time crystal, one which is astonishingly simple - a few particles suspended in a sound field which can organize themselves and keep moving in steady, repeating ways without being constantly driven. Instead of needing an external “push” to stay in motion, the particles interact through waves in a way that lets them draw in energy and convert it into collective behavior. What’s surprising is that this motion doesn’t come from the particles themselves, but from how they are arranged and how they interact.

In some cases, these interactions lead to continuous, self-sustained oscillations - essentially a system that keeps time on its own. In the simplest setup of just two particles, this becomes what’s called a classical time crystal: a stable rhythm that emerges naturally, rather than being imposed from outside. The exact frequency isn’t pre-set - it arises from the system itself. Small differences between particles also play an important role, helping stabilize these behaviors and allowing more complex dynamics in larger groups.

These findings point toward a different way of designing technology. Devices that can generate their own steady signals could become smaller and more energy-efficient, since they wouldn’t need constant external input. This could impact everything from electronics to medical devices. The same principles could also be used to build highly sensitive sensors, where stable, self-sustained motion makes it easier to detect tiny changes in the environment.

There are also implications for timekeeping, where naturally stable oscillations could provide reliable timing for systems like communications or navigation. More broadly, the work suggests a path toward self-organizing technologies—materials, robots, or systems that don’t need to be tightly controlled, but instead coordinate themselves through interactions. In that sense, it offers a new framework for building systems that are not just efficient, but also more autonomous and adaptive.

Community & other links

Phrase of the week: “Recursive self-improvement (RSI): AI systems are increasingly being used to build better AI systems, creating a feedback loop that could accelerate the very curves I showed you above”

Other Links:

• “To design AI for disruptive science, we would need to understand what “rules” make one paradigm better than another, and build systems that optimize for these. This turns out to be a harder problem than scaling compute. The answer cannot simply be experimental success, since experiments are slow and do not always reliably distinguish between paradigms (as was the case with Lorentz and Einstein). And there are other plausible candidates, but none yet offer a sufficient formulation.” -- Djajadikerta, Alvin. “Designing AI for Disruptive Science.” Asimov Press (2026). DOI: 10.62211/29ej-27et

• Google launches “Stitch”: Introducing “vibe design” with Stitch

• Anthropic started a new AI for Science blog

• NVIDIA and Global Robotics Leaders Take Physical AI to the Real World

• The Rise of AI-Powered Robotics: How 2026 Is Reshaping Manufacturing and Automation

Field Trip

Subscribe now

Did we miss anything? Would you like to contribute to Decoding Science by writing a guest post? Drop us a note here or chat with us on X.

Subscribe now

What we read

The 100x Research Institution [Andy Hall, Free Systems, Jan 2026] (HJ)

Andy Hall ran an experiment where he gave his 2020 vote-by-mail paper to Claude and asked it to replicate and extend the findings with new data. He then had a UCLA PhD student do the same work without AI assistance which provided a baseline to compare the results with. Claude’s output was very close – it correctly coded 29 of 30 California counties on treatment timing, and its collected data correlated above .999 with the manually gathered figures. Most importantly perhaps, despite the many mistakes, it finished within the hour, a task that otherwise took several days.

Hall uses this to build up to a broader institutional argument: “What would it mean to produce research with 100 times the scope of our vote-by-mail study? Not 100 times the length. 100 times the effort—because that’s what AI now makes plausible.”

Some applications are predictable – “more robustness checks, larger meta-analyses, deeper empirical exercises.” The more compelling idea is what he calls “living research” or empirical findings maintained as continuously updated infrastructure rather than static publications.

Hall’s other big vision culminates in a prototype he calls a “research swarm”: AI agents propose studies, an “LLM council” ranks them, agents execute the top candidates, and another LLM measures the output. Given how the economics of research have shifted, it is plausible to set-up this idea very cheaply.

There are almost certainly risks though and Hall flags three of them including a deluge of plausible-sounding but incorrect empirical work, incentive pressure toward small easily-verified studies crowding out ambitious ones, and massively scaled p-hacking as researchers can now search over huge spaces of possible analyses.

What we read

Material Science, Physics and Robotics (concrete advances outside of BioML):

Quo vadis biophotonics? Wearing serendipity and slow science as a badge of pride, and embracing biology [Schroeder-Turk, G.E., arXiv, Feb. 2026]

I will write about this article as the author wrote about science: from a first-person perspective and filled with passion, curiosity, and interest in the field of study. No LLM editing, raw.

Following on from last week’s theme of colour in biology, we are further diving into the field of biophotonics, this time with a funky opinionated article by G.E. Shroeder-Turk. Combining quotes from The Little Prince by Antoine de Saint-Exupéry with deep technical expertise on the field of biophotonics in the same texts, he argues for the need to understand biology in greater depth, beyond the ‘surface-level’ knowledge that suffices to allow for technological translation.

Consider the gyroid for example. A naturally occurring photonic crystal, a gyroid is what introduces the color green into some green-winged butterfly species. With a well-defined geometry, a gyroid can be fabricated synthetically in a bottom-up fashion using 3D printing, or in a top-down manner by templating nanostructures with polymers. Gyroids find use in optical sensors, photonic crystals, battery electrodes, bone tissue scaffolds, and filtration membranes - to name but a few applications. Effectively, the commercial use of the structure has not gone unnoticed. But what has - and what is a recurring pattern in the discovery and use of technology - is that commercial adoption often outpaces deep mechanistic understanding.

In the context of biophotonics, we can synthesize and replicate the biological curiosities nature has brought the field, but fail to understand the deep biological underpinnings of function and form. How does a butterfly wing grow? What mechanical stresses and consequential changes in gene expression do cells experience when forming a highly structured optically active geometry? And can we scaffold a mouse to grow butterfly wings - just as we tissue engineered it to grow a human ear? (1)

Three questions I was left wondering about post-article were:

1. Can we rethink how diagnostic sensors operate by having a direct conformational interaction with the material induce a color? (This question came from the reference discussed where direct interaction of a bacterium with silver induced a color change.)

2. What can we learn from how nature grows materials to either replicate the growth process, or use them directly by creating support structures to scale biology itself? (This question arose from the mention of using diatoms as ‘alternatives to cleanroom nanofabricated photonic crystals’.)

3. What other fields are similar to biophotonics in the sense of simply ‘stamp collecting’ new discoveries? Where do we leave data on the table that we could do more with, either by learning about the origin of how that ‘stamp’ came to be, or understanding the purpose and function of that ‘stamp’ in its wider biological context to be able to scale materials (towards the ultimate universal fabricator)?

I leave you to join the discussion in the comments, and close with the thought that biology is the deepest technology of all. So much more to discover.

Subscribe now

Did we miss anything? Would you like to contribute to Decoding Science by writing a guest post? Drop us a note here or chat with us on X.

Subscribe now

What we read

Computational design of blue melanin with peptide motif scaffolding [Lee, D.S. et al., biorXiv, Feb. 2026]

Ever since AlphaFold 3 came out, the question of when a new coloured protein with pre-designed band gap would be designed has been on my mind. Why coloured? Because this would indicate the orientation of amino acids has been tuned in such a way that its folding is controllable down to the nanoscale. Indeed, in green fluorescent protein (GFP), the most commonly used fluorescent protein for molecular biology, three amino acids of Ser65-Tyr66-Gly67 fold into a rigid conjugated structure. This structure in turn forms the chromatic core that emits in the green region of a light spectrum at 508 nm. Consequently, a single substitution from Tyr66 to histidine (Y66H) alters the local environment and causes the emission to shift from green to blue at 448 nm [1]. This BFP was already discovered in 1999, pointing to the fact that chromatic core mutation is not new.

So what is new about this recent work by Lee et al? The fact that colour was generated from ‘scratch’.

Melanin is a natural polymeric peptide mix that forms when the amino acid of tyrosine (Y) crosslinks into an extended coloured chain. This colour arises from a chain’s length; shorter chains absorb at shorter wavelengths - emitting red and yellow colours, whilst longer chains absorb longer wavelengths - emitting bluer colours. If a heterogenous mix of short and long peptides is present the overall polymerised products will emit a brown to black colour as a direct consequence of all spectra overlapping (figure 1a). Tuning colour thus requires tuning polymer length.

Realising this, Lee et al. set out to tune peptide chain extension in a controlled manner. Naturally, melanin polymerisation undergoes an enzymatic process. In this, a tyrosinase enzyme oxidises the amino acid of tyrosine (Y) to form the intermediate species of DOPAquinone. In the absence of a cysteine group, this species goes on to form DHI and DHICA polymers, both of which are classified as eumelanin. For the more chemistry-oriented readers a figure 1c represents the reaction in full.

Figure 1 | a) Schematic explanation of how natural melanin polymerises, yielding a heterogeneous mix of peptide lengths that together produce a black colour as visible at the macro scale. b) Introduction of a RFdiffusion designed polypeptide for controlled and narrow spectral absorption & emission tuning. c) Natural melanin formation pathway, starting from the tyrosine amino acid till DHI and DHICA polymeric products, as adapted from Cichorek M. et al. (2013) [2]. d) SPYYG peptide with a maximal absorption at max=617nm, creating a visible blue colour (what is not absorbed is what we see).

As may be seen from the mix of products formed, polymerisation in such a manner leads to a mix of brown and black. To circumvent heterogeneous product formation Lee et al modified the starting material to be a single pentapeptide (peptide with five amino acids) chain. By generating a library of 905 Y-containing peptides using RFdiffusion, 855 of which had one tyrosine, and 50 of which had 2 tyrosines. This library was subsequently screened for solubility, of which the top 10 possible designs were synthesised. From these a 20% hit rate resulted, with 2 out of 10 peptides forming a desired blue colour when polymerising up to a chain length of n=60 residues (figure 1d).

So what was so exciting about this result? Besides being able to retain physically beneficial properties of natural melanin - including high thermal stability (up to 121°C) and photostability, the authors were able to generatively produce a new, completely natural colour. Previous companies that have started purely as a function of being able to create natural pigments with nanoscale precision include Sparxell, Cypris Materials, and Colorifix. Whilst the biggest challenge for each of these remains being able to scale up nanoscale precise materials to the macro level and beyond - towards industrial-grade reproducibility, the introduction of a new way of fabricating materials in enzymatically compatible manners whilst still retaining nanoscale precision is a fascinating glimpse into the future of what other materials could emerge from this new world of protein design. Universal Fabricators thesis, here we (the materials engineers and protein designers of this world) come!

And sorry Avatar fans, blue skin is not on the horizon just yet; for that modification of melanin polymerisation in the complex environment of the human body the development of virtual cells, along with other biological tools, may be useful. Someday perhaps.

Integrated-photonics-based systems for polarization-gradient cooling of trapped ions [Sabrina M. Corsetti, et al., Light Sci Appl, Jan 2026]

Trapped-ion quantum systems offer strong prospects for quantum computation, but their performance and scalability are limited by cooling and optical control. Current experiments rely on free-space optical assemblies that are bulky, sensitive to vibration and drift, and difficult to scale as system complexity increases. As trapped-ion architectures move toward larger, integrated platforms, these optical constraints have become a central bottleneck.

This work demonstrates the first experimental realization of polarization-gradient cooling (PGC) of a trapped ion using an integrated-photonics platform. Unlike earlier integrated-photonics approaches focused on Doppler or resolved-sideband cooling, PGC enables faster cooling at lower optical power, directly improving operational efficiency. Implementing PGC on chip replaces complex free-space polarization optics with a compact, scalable alternative.

The approach is enabled by polarization-diverse integrated-photonics devices that generate intensity-matched beams with well-defined polarization states. Using a 200-mm wafer-scale silicon nitride/silicon dioxide process, the authors fabricated bilayer, apodized, curved grating emitters that produce unidirectional, focused beams compatible with planar ion traps. Three polarization configurations (TE-TE, TM-TM, and TE-TM) were explored and characterized through simulations and measurements of emission angle, beam shape, and polarization purity.

The authors also implemented single- and dual-fiber input systems, showing that on-chip splitting and routing provide passive phase stability for generating polarization gradients without active stabilization. Extensive testing across multiple chips and wafers demonstrated high device uniformity and reproducibility.

This platform enables scalable, phase-stable polarization-structured light fields and supports future extensions to multi-ion cooling, more complex photonic architectures, and advanced operations such as photon-mediated entanglement and electromagnetically induced transparency cooling. By integrating PGC into a chip-scale photonic system, the work addresses key limitations of free-space optics and advances scalable trapped-ion quantum hardware.

To summarize in layman terms, this work may be the beginning for scalable photonic chips!

Other

Apple has acquired Q.ai, a startup built by members of the original team behind Apple’s Face ID technology. Q.ai’s core research focuses on reading microexpressions and translating them into input signals for hardware like smart glasses, phones, and headsets. The thesis is straightforward: if you can reliably decode what the face is doing at millisecond resolution, you unlock an entirely new interaction modality that doesn’t require hands, voice, or even deliberate intent. Think scrolling by glancing, confirming by squinting, or dismissing a notification with a micro-frown — all without touching a screen or saying a word. The acquisition makes strategic sense. Apple has been steadily building toward a future where spatial computing and wearables replace the phone as the primary interface. Vision Pro already tracks eye movement; Q.ai’s microexpression models could extend that capability to the full musculature of the face, enabling far richer and more nuanced input than gaze direction alone.

Subscribe now

Did we miss anything? Would you like to contribute to Decoding Science by writing a guest post? Drop us a note here or chat with us on X.

Subscribe now

We are still looking for new members. If you’re excited not just about reading about breakthroughs at the intersection of AI and science, but about shaping the narrative around them, we want to hear from you. Drop your details here.

What we read

Learning to Discover at Test Time [Yuksekgonul M. et al., arXiv, Jan. 2026]

Can we iterate our way towards new ideas?

The space beyond the present state-of-the-art is where patents are filed and progress is made. Defining this space however is practically impossible, as this task would require defining the space of infinite possibilities that extrapolate from the current cutting-edge. But what if there was a way to work around this; what if instead of solving for all possible future solutions, we could create a method that narrows in and focuses on improving only the state-of-the-art for a given prompt?

Still, iterating beyond what is already known in a rational and progress-driven manner is no easy feat. In fact this is what most scientific labs, industry, and PhD students spend the majority of their time working on. And yet, Yuksekgonul M. et al. demonstrated it was possible to surpass this current ‘best’ for a given prompt.

So how did they do it? First, they took classic Test-Time Training (TTT) [1] and adapted it to add two unique properties: 1. an entropic maximisation objective, and 2. a reuse rule. In TTT a policy continues to learn at inference as the model is updated to focus on the specific problem at hand. The TTT goal is to find one state that beats the current state-of-the-art. This contrasts with standard RL, where the goal is to learn a policy that will be deployed repeatedly and thus must perform well on average, generalising to the state distribution induced by environment* dynamics + the initial state distribution.

Figure 1 | Comparison of the TTT-Discovery method implemented by Yuksekgonul M. et al. as compared to Best-of-N approach.

Introducing an entropic maximisation objective pushed updates towards the highest-reward trajectories - emulating how one might iteratively learn their way towards the best outcome. The second modification, of introducing a reuse rule, required the starting state of the new iteration to be sampled from an archive of previous attempts. Using a reuse rule had two advantages. First, because the initiation state was not always restricted to the state-of-the-art result the model could explore a wider range of possibilities. Contextualising to the world of human ideas we see breakthroughs continue to be made across domains because new individuals continue to enter, each joining a field with a slightly different bias (their background). As a result, the types of questions they will ask also differ. This is analogous to how the model started at a new state each time. The second advantage of initiating the model with previous states was that the reinforcement learning policy could explore with greater depth, as it did not need to start from scratch each time. Effectively, one can think of this as a student going to a lecture, and then learning that content to take it forward, thinking deeper about the specific task at hand.

Ultimately what makes TTT-Discover exciting is the ability to go beyond searching for a good solution: it searches for a new solution; one that surpasses the state-of-the-art. Defining the space of outcomes by leaning into rare and high-reward breakthrough attempts, whilst also ensuring the model is not reset to 0 each time, yielded results in single-cell sequencing (biology), algorithm engineering (mathematics), and kernel engineering (coding). In the context of IP, industry, and development it will be interesting to track how this accelerates invention, closed-loop discovery models, and shifts how inventions are valued.

*The nuric Blog gives a great explanation on Markov Decision Processes and how environments are defined.

The Adolescence of Technology [Dario Amodei, Jan. 2026]

Dario Amodei, the founder of Anthropic, has followed contrary to his optimistic vision of “Machines of Loving Grace” with a cautionary essay titled “The Adolescence of Technology.” This work serves as a stark counterpoint, shifting focus from a utopian future toward the catastrophic risks of what he calls “Powerful AI”. Amodei estimates so-called “Powerful AI,” which could be 1–2 years away or far longer, will come rather sooner than later. He argues this is due to automation of coding, novel math discoveries, and the accelerated technological advances Dario already sees at Anthropic.

He then starts the main thought experiment: what would you be worried about in a country of geniuses, 50 million people with higher intellect than Nobel Prize winners and the biggest innovators? He categorizes the risks in five brackets:

The first is Autonomy Risk, the concern that a superintelligent system might seek power to ensure its objectives are met. While some argue models are merely trained to follow instructions, Amodei notes that as models become more psychologically complex, trained on vast literatures describing power dynamics. They may naturally develop power-seeking traits. To defend against this, he suggests a combination of rigorous technical alignment, mechanistic interpretability to peer into the model’s “thoughts”, and creating a system of checks and balances where different AI models monitor one another.

The second risk is Misuse for Destruction, where powerful models enable non-state actors or individuals to develop biological weapons or orchestrate massive cyberattacks. The primary defense here are screening systems that detect dangerous queries, restricting model access in high-risk domains, for instance coordinating with biosecurity agencies, and building detection capabilities for novel pathogens.

The third risk, Misuse for Seizing Power, involves nation-states leveraging AI to establish absolute military or surveillance dominance. Amodei argues that the defense must be geopolitical. Democratic nations must maintain a lead in AI development to ensure that the “intelligence balance of power” remains in favor of open societies, potentially through international coalitions that manage compute resources.

The fourth risk is Economic Disruption, as labor markets are upended at a pace faster than human institutions can adapt. His defensive suggestion focuses on a radical restructuring of the social contract, potentially involving Universal Basic Income or human-centric industrial policies that ensure the dividends of AI-driven productivity are shared broadly rather than concentrated in the hands of a few.

Finally, he warns of Indirect Effects on the human condition, ranging from biological shifts to a fundamental loss of human purpose as machines outperform us in every creative and intellectual endeavor. The defense for this is more philosophical: we must deliberately choose to keep humans in the loop for meaningful tasks, ensuring that technology enhances rather than replaces the human experience.

Amodei concludes that this transition is a humanity-scale test. Our success depends on whether we can build these robust defenses before the digital geniuses outpace our ability to manage them.

RL Environments and RL for Science: Data Foundries and Multi-Agent Architectures [Kourabi & Patel, semianalysis, Jan. 2026]

The authors describe the upcoming influence of RL in AI. They argue, the gold standard for model improvement was always more compute during the pretraining phase, but the last 18 months have signaled a change led by OpenAI’s efforts to scale up post-training reinforcement learning. While other frontier models like Google’s Gemini still rely primarily on massive pretraining, OpenAI has moved toward a paradigm where the model continues to learn through interaction. To evaluate these leaps in performance, they created GDPVal, a framework designed to test how well these RL models actually generalize.

The primary challenge in this new field remains the difficulty of training; unlike pretraining which consumes the static internet, RL requires building complex data and tasks from scratch. This bottleneck has created its own startup ecosystem focused on environment synthesis, ranging from cloning websites (like doordash) and mimicking software interfaces (like salesforce) to maintaining massive, sandboxed coding environments where models can execute and iterate on code in real-time.

The authors argue that we are entering a phase defined by data, specifically the kind of interactive data that allows for RL-as a-Service. Startups offer tailored RL to enterprises, often using Qwen models that are easy to post-train and OpenAI launched their own service mainly aimed at building customized services for enterprise users.

According to the authors, the biggest opportunity lies in RL for Science, where the focus shifts from language modeling to hypothesis testing. Periodic Labs emerged with a $300 million seed to build an AI scientist using closed-loop RL grounded in physical experiments. Models propose hypotheses, test them in simulators, then inform real lab work – mapping roughly onto a graduate student’s workflow. Physical RL is uniquely hard. Biology experiments take days and cost thousands of dollars, compared to coding tasks that can run 64 times for trivial cost. Scientific literature often disagrees even on basic questions. Companies are building robotically automated labs not to design drugs, but to generate the validated data that foundation models need.

RL environments are spilling into the physical world. They’re no longer Docker containers but real experiments with real costs. The future moat for AI development is no longer just the size of the cluster, but the sophistication of the environments where these models learn to think.

Field Trip

Subscribe now

Did we miss anything? Would you like to contribute to Decoding Science by writing a guest post? Drop us a note here or chat with us on X.

Subscribe now

As we kick off 2026 we are more than excited to expand our team and gain some fresh perspectives. If you're excited not just about reading about breakthroughs at the intersection of AI and science, but about shaping the narrative around them, we want to hear from you. Drop your details here.

What we read

Conditional Memory via Scalable Lookup: A New Axis of Sparsity for Large Language Models [Cheng X. et al., arXiv, Jan. 2026]

When constructing a sentence your brain does not learn each word nor sentence structure from scratch. What is novel about the sentence in formation however is the context it is responding to. In the modern English language for example the top 10 most used words are: ‘the’, ‘be’, ‘to’, ‘of’, ‘and’, ‘a’, ‘in’, ‘that’, ‘have’, ‘I’ [1]. The context in which these are used is commonly the same; the will most probably come before a noun, ‘I’ most often is followed by a verb… and so forth. If compositional reasoning, which is what transformers enable, can be thought of as how words are combined to construct a sentence, then knowledge retrieval is the knowing of the existence of these words, and which are chosen most often.

Mixture-of-Experts (MoE) has become the default for scaling frontier LLMs; it allows dividing compute across specialist subnetworks, making the model use only a fraction of the model per token rather than activating the entire net. As a result parameter count can be inflated without inflating the cost per token. However, whilst MoE reduces the cost per token it cannot bypass the need to recompute ‘known’ facts. As a result, today’s models often retrieve reconstructing static patterns through layer-by-layer computation spending early-layer attention re-deriving things that are very similar to lookup tables.

Here, Xin Cheng and the team at Deepseek proposed the use of a second type of memory to complement model input → output transformations. For this, an Engram conditional memory module was introduced (figure 1). Engram operated in two phases: retrieval and fusion. First, it takes suffix N-grams from the token stream, converting these into hashes with deterministic IDs that can be recalled from precomputed lookup tables. Using an All-to-All approach, the Engram table was distributed across GPUs; because GPUs can send and receive data from one another, this effectively allowed scaling beyond a single GPU, storing the full Engram table in memory. It also meant compute occurred locally and only required embeddings to move around.

Figure 1 | Engram conditional memory module architecture. To the left (a) is a zoom-in of how an Engram module operates. N-gram static vectors, here shown as 2-gram or 3-gram, are transformed into hashes that match a precomputed ‘lookup table’. Concatenation of resultant matching vectors is followed by gating vectors with the current state to ensure only relevant memory passes through, and finally data is smoothed over with a convolutional layer. During training (a) the lookup table is split across GPUs using an All-to-All approach. At inference time (b) the Engram table is offloaded to host CPU memory, and only rows necessary to respond to an input query are fetched from this ahead of time, such that at the GPU compute time overhead stays small. In this manner, the GPU can ‘focus’ on computing other layers rather than stalling to fetch a massive memory table into VRAM or waiting on synchronous data transfers, as would be the case for other transformer models.

Beyond the Engram module, the critical contribution of the paper is formalisation of the Sparsity allocation problem, which asks the question of how, under fixed total parameters and fixed activated compute, sparse capacity should be split between MoE experts (conditional compute) and Engram tables (conditional memory)? Investigating this balance, the authors identified a U-shaped relationship, where too much MoE wastes compute on lookup-like work, whilst too much Engram induces losses in dynamic computation required for true reasoning. As a result, performance was optimal when ~20–25% of the sparse parameter budget was reallocated from MoE to Engram.

Scaling thus becomes a balance of how much compute can be sparsified using MoE, as well as how much memory can be sparsified with Engram modules. How a model allocates attention, compute, and lookup behaviour will be the focus of the next research areas: going back to the English language analogy - we need poets to write Shakespeare, editors to recall the rules of grammar, and lexicographers to invent new words. So write, and code, and write!

What we read

Scientific Production in the Era of Large Language Models [Kusumegi et al., Science, Dec 2025]

In this paper, Kusumegi et al. use text-based detection to analyze the abstracts of about 2 million preprints to see what happens to science as a system when LLMs enter the production process. An author is marked an “adopter” of these models at the first manuscript where statistical signatures of LLM assistance cross a detection threshold. Then, by comparing submission rates between adopters and matched non-adopters, they find that LLM use correlates with productivity jumps of 36–60% depending on repository (arXiv, bioRxiv, and SSRN) with the largest gains accruing to non-native English speakers: “A productivity jump may stem from the use of Gen AI across multiple research tasks, including idea generation, literature discovery, coding, data collection, or analysis. But to date, LLMs likely have had the largest impact in writing.”

This is consequential because their second finding focuses on the relationship between written complexity and signals of quality. In the past, “writing complexity was positively associated with manuscript quality, as approximated by the probability of publication in a peer-reviewed venue.” Yet, despite LLM assisted preprints having higher language complexity, they were seen as less meritorious – more sophisticated language predicted worse peer assessments.

However, manuscripts in the “adopter” group did benefit from more diverse literature discovery: “We compared publication formats, showing that Bing users access books at a 26.3% higher rate [relative to accesses to arXiv manuscripts redirected from Google]...; the median age of manuscripts accessed decreased by an estimated 0.18 years.” Citation behaviour also followed similar patterns.

Although the analysis has its limitations, including failures to identify human-edited LLM text, there are real policy concerns. As writing quality becomes uninformative while output volume surges, “editors and reviewers may increasingly rely on status markers such as author pedigree and institutional affiliation as signals of quality, ironically counteracting the democratizing effects of LLMs on scientific production.”

What we read

Surface-engineered anisotropic Fe3O4 nanoplates for highly efficient magnetic field-assisted micro/nanoplastics remediation [Jeong et al., Journal of Environmental Management, Jan 2026]

Micro- and nanoplastics have become some of the most persistent pollutants of modern life. They are now found everywhere from oceans to human bodies, resisting natural degradation and raising growing health concerns. Existing removal strategies struggle to keep pace: filtration and coagulation are costly and prone to fouling, while biological approaches are often too slow. As the scale of the problem becomes clearer, the need for fast, efficient, and deployable remediation technologies has become urgent.

Magnetite (Fe₃O₄) has long been valued for its strong magnetic properties, making magnetic nanoparticles an attractive option for plastic removal. However, conventional spherical designs often lose efficiency over repeated cycles and offer limited adsorption due to poorly tailored surfaces. To overcome these limitations, Jeong et al. moved beyond isotropic particles, developing anisotropic magnetite nanoplates with engineered surface chemistries. By combining particle shape, functional interfaces, and magnetic-field-assisted motion, they aimed to create a more effective nanoharvester for micro- and nanoplastics.

The resulting silica-coated magnetite nanoplates, further functionalized with aminopropyl, octadecyl, or phenylethyl groups, showed striking performance. Under an applied magnetic field, the optimally functionalized MNPL@SiO₂ removed over 92–94% of polystyrene micro- and nanoplastics ranging from 100 nm to 1 µm within 10 minutes. Kinetic analysis followed a pseudo-second-order model consistent with chemisorption driven by electrostatic interactions, while equilibrium behavior fit a Langmuir model with a high sorption capacity of ~3630 mg/g.

Notably, adsorption was only part of the mechanism. Under a magnetic field, the plate-like nanoplates assembled into hierarchical structures that physically trapped additional plastic particles—a process termed dynamic trapping—which contributed an additional ~18% removal efficiency. The amine-functionalized nanoplates also retained strong performance over multiple reuse cycles after solvent cleaning, highlighting their potential for scalable deployment.

By moving beyond classical adsorption and introducing dynamic trapping, this work outlines a scalable and efficient route toward rapid water remediation, particularly for nanoscale plastics that have proven hardest to remove. The approach shows clear promise for wastewater treatment, industrial effluents, and broader environmental cleanup efforts.

Community & other links

Lyte emerges from stealth, building a 3D sensing platform for robotics and Physical AI

Cambium raises $100M to develop its AI-driven Material Discovery engine.

Nvidia and Thermo Fisher Scientific partner to increase the automation and speed of lab instruments.

Subscribe now

Did we miss anything? Would you like to contribute to Decoding Science by writing a guest post? Drop us a note here or chat with us on X.

Subscribe now

What we read

Please, don’t automate science! [Julian Togelius, December 2025]

Panelists at an AI for Science event at NeurIPS discussed plans to replace humans at all levels in the scientific process. Togelius argues that fully automating science (strong science automation) would be a profound mistake as it would strip humans from one of the deepest sources of meaning which is contributing to the growth of knowledge. He argues that even a faster cure to cancer in exchange for removing humans out of the loop in science is not worth it. Scientific progress guided by humans, in Togelius’ mind, is more valuable than accelerating outcomes.

Weak science automation, where AI acts as a tool that enhances human productivity and creativity, would be beneficial as it enhances human intellectual and creative work rather than making it obsolete. He also argues that if science experiences strong automation, merely making a hobby for humans, it would be a pretend version of advancing knowledge, not actually doing it.

Togelius says that full automation of science is not inevitable, and that we should be at the wheel making decisions; it is not an unstoppable train but a truck whose direction society can choose to steer.

Context Windows [Kevin Baker, Artificial Bureaucracy, Dec 2025]

Baker argues that the debate over whether LLMs can “do science” over-indexes on the “capabilities question.” Instead, we should be focusing on whether these models gel well with the scientific program that “was already running when they arrived.”

As many have pointed out, the current way of doing and sharing science is quite dysfunctional. Here, Baker traces this problem back to the 1960s, when we started quantifying science via citation metrics. These numbers were originally built to help researchers navigate exploding literature volumes but (predictably) morphed into signals of credit and endorsement. This represents what Robert Merton called “goal displacement,” the process by which the means become the ends.

The usual response to this issue invokes Goodhart’s Law: fix the corrupted measure with better measures. But Baker contends that this framing is insufficient. Goodhart says find a more robust indicator while Merton says the indicator is doing precisely what system design intends. Chasing new metrics ignores how measurement itself warps the relationship between good research and prestige.

With this context, we can appreciate how LLMs worsen the current dynamic by accelerating the “optimization machine.” We can now produce more stuff faster, the “paper-mill artifacts” that pollute the scientific commons. Baker notes that “one might hope that this acceleration heightens the contradictions, that the systems produce so much slop so quickly that the problem finally becomes undeniable.” But to truly avoid persisting in this state of dysfunction, we need “those who can build countervailing power, and who decide to change what gets measured, or finally wrench the institution of science itself from the false promise of measurement.”

Also see: Ben Recht’s response to Baker in Measures as Ends.

What we read

Combinatorial asymmetric acoustic metamaterials with real-time programmability [Melanie R. Keogh and Osama R. Bilal, Oct. 2025] - SC

Sound control, much like heat management in electronics, often runs into a hard physical limitation: once acoustic metamaterials are fabricated, their geometry—and therefore their function—is fixed. This has long constrained their ability to adapt in real time. While tunable designs exist, achieving scalable, on-the-fly programmability has remained a major challenge.

Researchers at the University of Connecticut have now demonstrated a way around this, with a real-time programmable acoustic metamaterial built from asymmetric pillars with indentations that can rotate freely in the plane at any angle. Watch the video here shared by UConn! Instead of redesigning or refabricating the structure, the behavior of the material can be changed simply by rotating its components.

The impact of this approach is its sheer flexibility. Even a single unit cell, adjustable in 1° increments, yields 180 distinct configurations. When scaled to an array, this explodes into a design space exceeding 10¹⁰⁰ possible combinations—far beyond what conventional, static metamaterials can offer.

The team validated the concept through simulations and experiments using an 11×11 array of reprogrammable pillars. By rotating the pillars to different orientations, they demonstrated precise control over wave propagation: at 0° the structure strongly attenuated sound at predicted band-gap frequencies, while at 90° it allowed full transmission. They also showed topological insulator behavior by setting all pillars to 45°, forcing sound waves at 12.46 kHz to travel only along the edges without backscattering. Finally, by combining multiple orientation domains (0°, 45°, and 135°), they achieved diode-like, nonreciprocal sound propagation—blocking waves from one direction while allowing them from the other.

Together, these results highlight a shift from static acoustic materials to programmable platforms that can respond dynamically to their environment, opening the door to a new generation of adaptive sound-control devices.

Community & other links

SRPIN-D, the German innovation agency, is launching the “Next Frontier AI” initiative. They will support 10 teams for 2 years with up to EUR 125M in funding, and three of the teams have the potential to receive EUR 1B. The aim is to create European frontier AI labs.

Excelsior Sciences raised $95M to advance its AI-based chemical synthesis platform for small-molecule drugs.

Unconventional AI emerges from stealth with a $475M seed round. They develop energy-efficient computers for AI. Their architectural design takes inspiration from biology and the brain.

Medra secures $52M to develop physical AI scientists for drug discovery.

Field Trip

Subscribe now

Did we miss anything? Would you like to contribute to Decoding Science by writing a guest post? Drop us a note here or chat with us on X.

Subscribe now

What we read

Latent Collaboration in Multi-Agent Systems [Jiaru Zou, arXiv, Nov. 2025]

Have you ever had a thought that makes perfect sense but cannot be explained in words? This is what last-layer hidden embeddings in a large language model are like.

In a neural network - a large language model of which is a particular subtype, input sequences are encoded using an embedding layer. During the embedding process each token (a word or subword) is mapped via a learnable embedding matrix Win to a dense embedding vector e. Because all embedding vectors have the same size they can be stacked together into a single vector E, which can pass through the model layers. Each layer has an associated hidden state; the last layer of neurons prior to decoding back into text tokens thus contains the last-layer hidden embeddings.

Previous research exploring the use of ‘model language’ to allow multi-agent systems (MAS) to communicate with one another has proposed layer embeddings for direct `discussion’ across models. Yet what Zou et al. introduce with LatentMAS is a fully-latent-space system, where both reasoning and communication occur in the latent space. As shown in figure 1, two key interventions make this possible. First, a key-value cache is created for storing last-layer hidden embeddings (ht). Effectively, the key-value cache may be thought of as a large library of input tokens (keys) that map to last-layer hidden embeddings (values). Each time a token is encoded this key-value (KV) pair is added to the key and value vectors, creating an extensive lookup table that stores all previous mappings. In LLMs using a KV cache saves recomputing attention over the entire input prompt (prefix), making generation of tokens linear in computation required. Second, the KV cache of each agent is shared with all other agents. This means that if a KV pair already exists, another agent may benefit from its encoded ‘knowledge’ even though it was not originally an expert in this space. Effectively, the latent working memory allows math models in LatentMAS to interact with science and code models without requiring the decoding of each token back into text. As a result LatentMAS was demonstrated 4-4.3x faster than text-base MAS at end-to-end inference, with the former achieving performance similar to the latter despite using only <50 inference steps only as opposed to the >20,000 steps TextMAS required.

But did these last-layer hidden embeddings indeed accurately represent text-based outcomes, and were results still reliable after a large number of inferences? For the former, yes. Provided each output last-layer hidden embedding ht+1 was correctly realigned with input embeddings using projection matrix Wa, Zou et al observed tokens were closely aligned with those of TextMAS when projected into a separate latent space for comparison. For the latter, the answer depends, with optimal performance of LatentMAS occurring at 40-80 steps - beyond which performance plateaued or even dropped slightly.

The biggest takeaway: multi-agent systems may communicate in lossless form and without any additional training required to convert from one ‘output’ to the next by shifting the entire discussion into the latent space. And in terms of the thought in your head that is just too difficult to explain? Perhaps we must wait till our latent thoughts can be directly vectorised and projected into a shared latent space… brain chip anyone?

Figure 1 | Overview of the multi-agent system (MAS) implemented. Agents 1-N are labeled A1, A2, …, AN in the graphic, with N=9 in LatentMAS. Input sequences follow a conventional transformer embedding process, yet unlike text-based MAS output tokens remain in the latent space, with last-layer hidden states used for subsequent model input training. Key-value (KV) cached pairs are shared between agents in a latent working memory. Previous KV values from agent A1 are also embedded directly into subsequent agent inputs (arrow from latent working memory of A1 to A2) to speed up computation and lower costs.

CODEDISTILLER: Automatically Generating Code Libraries for Scientific Coding Agents [Jansen et al., arXiv, November 2025]

Automated Scientific Discovery (ASD) systems (as named in this paper) can generate and run code-based experiments but are limited by the code they can reliably generate from parametric knowledge alone; knowledge that is learned at training time. Scientific, code-based experiments, however, require specific methods, measurements and protocols. To aid ASDs in this task the experimenter can allow the agent to “mutate” existing experiment code, or provide a library of existing code examples agents can combine.

CodeDistiller is a system that automatically builds large collections of scientific GitHub repositories into a library of domain-specific code examples. The system automatically converts GitHub repositories in specialized scientific domains into debugged, working examples that can be incorporated into ASD systems. The best base model achieved 74% performance on the distillation task in the materials science domain, which improved downstream experiment quality

What we read

WeatherNext 2: Our Most Advanced Weather Forecasting Model [WeatherNext Team, Google DeepMind, Nov 2025]

Google DeepMind has released an upgraded version of their AI weather forecasting system. The new model significantly improves on its predecessor in both speed and precision: it runs 8x faster and can make predictions at finer time intervals (hourly rather than longer windows). Perhaps most notably, it can generate a wide range of potential weather outcomes from a single set of starting conditions.

This is made possible by Functional Generative Networks (FGN), a technique that allows for the introduction of noise into the model in a way that keeps predictions “physically realistic.” FGN is also interesting because the model is trained only on “marginals” or discrete, individual weather elements at a particular location (temperature at one spot, wind at another). Even so, it learns to predict the “joints” or how these elements interact across “large, complex, interconnected systems.” The gains from this approach are substantial. WeatherNext 2 reportedly outperforms its predecessor on “99.9% of variables and lead times.”

Community & other links

Black Forest Labs, founded by pioneers in latent diffusion, raised $300M to rival image generation models from Google, ByteDance, and OpenAI. Their bigger picture is laying the foundation for visual intelligence.

Field Trip

Subscribe now

Did we miss anything? Would you like to contribute to Decoding Science by writing a guest post? Drop us a note here or chat with us on X.

Subscribe now

We had the pleasure of speaking with Antony Rowstron and Aayush Chadha from the Advanced Research + Invention Agency (ARIA), UK’s R&D funding agency built to unlock scientific and technological breakthroughs.

ARIA recently announced its funding programme for AI Scientist systems; autonomous systems that can generate hypotheses, design and run experiments, and analyze results with minimal human oversight. AI scientists have been discussed at length this year (some great write ups here and here), and we were keen to understand how ARIA believes AI scientists will lead to novel breakthroughs in science.

In this interview we discuss:

• ARIA’s AI Scientist call as a way to cut through noise, map what these systems can truly do today, and identify teams whose AI scientists can make real progress on difficult, interdisciplinary problems such as mitochondrial gene delivery.

• How AI scientists will accelerate human research rather than replace scientists, enabling many more “shots on goal” via parallel experimentation.

• How the UK will maintain scientific capability through advanced automation and democratized access to experimentation.

Antony Rowstron became ARIA’s inaugural CTO in June 2025 after 26 years at Microsoft Research. He’s led multidisciplinary teams working on a diverse set of subjects: optical storage technologies to robotics for data centres. Aayush Chadha is a Frontier Specialist, focusing on how to accelerate research in labs across the UK. He’s worked on a diverse range of scientific disciplines, from ML research to material production and scaling.

Pablo: Could you introduce yourselves and describe your roles at ARIA?

Ant:

I’m the CTO at ARIA. I joined about six months ago after 26 years at Microsoft Research, where I led research across robotics, hardware, and new device development. I’m a computer scientist by training, not a domain scientist, but my role at ARIA spans several areas. I mentor our frontier specialists and I support program directors as they define and refine new program ideas.

Another part of my role is to help ARIA embrace AI properly. Internally that means improving the tools and workflows we use. Strategically it means understanding how AI can increase the velocity of scientific discovery across our portfolio, and how we can push AI in directions that make it more useful for science rather than just consuming whatever industry produces.

I am particularly focused on the emerging “AI scientist” space: systems that can generate hypotheses, design experiments, run them via automated labs, analyze results, and iterate. Several companies are starting to build things in this direction. A major part of my job is to understand how ARIA should interact with such systems and how our creators might collaborate with them.

Aayush:

I started out in computer science, working on early deep learning systems such as recurrent neural networks and LSTMs. I later moved into materials science, where I worked on graphene and quantum dots and dealt with a lot of material scaling challenges. After that I spent some time at Entrepreneur First trying to start a company, and then I joined ARIA.

At ARIA I work across multiple programs, mainly neurotechnology, manufacturing abundance, and our AI scientist efforts. Recently I have been visiting many labs in the UK, especially in materials science and chemistry, to understand their current level of automation and where there are opportunities to accelerate research.

Ant, how does your previous experience, especially at Microsoft, shape your work at ARIA?

Ant:

At Microsoft I was deeply involved in building early cloud infrastructure. That gave me a strong sense of how you design very large technical systems for reliability, scale, and manageability. You think hard about uptime, standardization, abstraction layers, and how to make complex hardware behave like a dependable utility.

When I look at today’s labs, including the attempts at “cloud labs,” I rarely see that kind of systems thinking. Labs are often fragile, bespoke, and heavily dependent on people physically being there at odd hours. The instruments are usually designed for human operators and human interpretation.

Bringing cloud-style engineering into the lab world could have a huge impact. You want laboratory infrastructure that runs 24/7, 365 days a year, with high availability and minimal manual intervention. For AI scientists to be effective, the physical layer has to be that reliable. Without that foundation, even very capable AI systems will be constrained.

That experience also shapes how I think about where ARIA can add value. We are not just funding new algorithms. We are thinking about how computation, instruments, and automation come together to make a step-change in how science is done.

What is ARIA’s objective with the AI Scientist call, and what would success look like?

Ant:

The call is our first structured way to engage with AI scientist systems. We are trying to understand what exists today, what these systems can actually do, and where their limits are.

We asked applicants to show us one problem they believe their system can solve now and a stretch problem they think it cannot yet solve. That gives us a picture of the claimed frontier and where they themselves see the boundary. It also helps us distinguish between marketing and reality, which is always necessary in a hype-heavy area.

We expect many submissions to be narrow and domain specific. That is completely fine. At this stage, crawling is meaningful: a system that can reliably execute a well-defined scientific loop is already impressive. At the same time, we hope to see a few attempts at more general or more flexible systems.

Success for this call would mean several things. First, we gain a realistic map of current capabilities. Second, we learn how to evaluate and interact with these systems. Third, we identify teams and approaches that could plug into future ARIA programs as true participants in scientific work rather than just as tools on the side.

Aayush:

Some of the problems in the call are closely tied to our existing opportunity spaces and programs. A very concrete success would be a team whose AI scientist can make genuine progress on one of those problems, earlier than we expected a human-led team to do so.

Today’s AI scientist systems that are publicly described are mostly very narrow. ARIA’s programs tend to be interdisciplinary. Demonstrating that an AI scientist can bridge at least two substantial domains would be a strong signal that the field is maturing.

Could you give an example of the kinds of problems you included in the call?

Aayush:

One example is from our program on engineering mitochondria. The goal is to deliver nucleic acids into the mitochondrial matrix. That has been a long-standing challenge in biology. Many groups have worked on it for decades without a fully satisfactory solution.

However, there is now an inflection point. Advances in nanoparticle-based delivery and related physical chemistry suggest new routes to the problem. A useful AI scientist in this context would need to understand mitochondrial biology and the complexes on the mitochondrial membrane, but also be able to draw on knowledge from nanoparticle design and materials science.

A system that can connect those areas and explore design spaces for delivery vehicles, identify candidates that are more likely to be taken up, and design experiments to test them would be very valuable. It is a good example of the interdisciplinary capability we are interested in.

Subscribe now

Are you mainly seeing specialized systems, like a CRISPR-GPT, or are you expecting general AI scientists?

Ant:

We’re working our way through the applications now, but they’re generally specialized. That is what the current ecosystem is geared toward. My hope is that at least a small proportion will be more general in scope.

What is already clear is that there is a lot of interest. We received hundreds of applications, far more than a typical call. That is exciting and slightly daunting, but it is a good sign of how much energy there is in this space.

I am not concerned if most of what we see is still at the “crawling” stage. This is a technically difficult area. If it were easy, it would not be appropriate for ARIA. Direction, credible progress, and clear thinking about limitations are more important than grand claims at this point.

Q: What happens after this initial call? Will teams be able to apply these AI scientists to full ARIA programs?

Ant:

Many applicants have already asked how they can participate in our main programs. The answer is that this call is partly about learning how to make that possible.

There are at least two layers to think about. One is how AI scientists participate as “creators” in programs, alongside human-led teams. The other is the underlying lab infrastructure. If we want AI scientists to plan and run experiments at scale, labs need to look very different to how they look today.

My experience with cloud infrastructure is relevant here. In the cloud, you hide a huge amount of complexity behind stable interfaces and you build for high availability. When I look at lab automation today, including attempts at cloud labs, I do not think we are yet at that stage of maturity. There is a lot of room for ARIA to support work on more robust, modular and highly available lab systems.

Aayush:

This is also where instrument design comes in. Many instruments are optimized for human operators and for human-readable outputs. An MRI scanner is a good example. The raw data are nothing like the image a radiologist sees. There is a large processing pipeline to turn a very complex signal into something a human can interpret visually.

If you free yourself from the constraint that outputs must be directly interpretable by humans, you can design instruments that produce data in forms that are much more natural for machine learning systems. CERN is an extreme example in physics, where the raw data are far too complex for humans to look at directly. We think there is room for that kind of thinking in many more areas of science.

Do you view AI scientists as something that should be sovereign, perhaps in a national capability or national security sense?

Ant:

ARIA is not a defense agency. We are the non-defense counterpart of something like DARPA. So we do not frame our work as a national security project.

At the same time, it is obviously important that the UK remains a serious player as science becomes more automated. The UK is world leading in many areas of science. To stay that way, our researchers need access to the right facilities, including advanced automation and AI-driven infrastructure.

My focus is less on sovereignty as a security issue and more on capability. We want UK scientists to have state-of-the-art tools, including AI scientists and the labs they require, so that they can continue to do world-class work.

How do you expect the role of human scientists to change over the next five to ten years?

Ant:

I think in terms of velocity. AI will increase the speed at which humans can generate and test ideas. That is similar to what the cloud did for computing. It did not eliminate programmers, but it changed what they could do in a given amount of time.

For at least the next decade, I expect AI systems to be tools that augment human scientists rather than independent agents replacing them. The history of technology supports that view. Look at autonomous vehicles. The hype always runs ahead of reality. Autonomous driving programs started in the early 2000s and only now are we seeing limited deployments of robo-taxis in a few cities. Human drivers have not disappeared. The same pattern is likely here.

So I expect AI scientists to become very powerful partners, but not to remove humans from the loop any time soon. Humans will still be central in setting directions, judging what matters, and interpreting the broader implications of results.

Aayush:

Higher velocity also changes the structure of research. In biology and other sciences we often talk about “multiple shots on goal,” but historically the number of shots you can take is limited by experimental bandwidth. Automated experimentation and AI-driven design allow a lead scientist to pursue many hypotheses in parallel.

You can imagine a future where a principal investigator defines a research quest and an AI scientist explores dozens of branches at once, generates results, and suggests the most promising paths. That kind of parallelism is a step-change, not a five percent improvement.

Do you worry that AI scientists could constrain scientific creativity or homogenize scientific direction? How do you think about that risk?

Ant:

Technological waves will happen whether we like it or not. The question is whether you shape them or are shaped by them. I have lived through several smaller waves in computing. The pattern is always the same. You either ride and steer the wave, or you get overwhelmed by it.

ARIA’s responsibility is to participate in this wave in a way that keeps it aligned with what is good for science and society. As a funding body we can influence which approaches get scaled. We can insist on human oversight where it matters, encourage diversity of approaches and make sure the incentives are not set up in a way that narrows science prematurely.

We are also investing a lot of thought in ethics and governance inside programs. It is not just about what is technically possible. It is about what kind of scientific ecosystem we want to build as these tools mature.

What does your ideal lab look like in five to ten years?

Ant:

There are two main aspects. The first is interdisciplinary instrumentation. Today labs are siloed. Biology instruments live in biology labs, materials instruments in materials labs, and so on. In my experience, many breakthroughs happen at the interface between fields.

I would love to see labs where unusual combinations of instruments sit side by side and are easily accessible to researchers and to AI systems. Ideally we would see discoveries that the scientists involved can honestly say would not have occurred if those tools had not been colocated, or if a system had not suggested an unexpected combination.

The second is full automation. When I was at Microsoft, our team took optical experiments that used to involve physicists turning knobs and logging values by hand and turned them into fully automated rigs. We added motorized stages, sensors and control software, then streamed all the data. At one point we were sending more data into UK data centers than any other UK customer. The scientists went from standing at a bench all day to writing Python scripts, pressing run and getting graphs a few hours later. The speedup was dramatic.

I recently heard about students staying until four in the morning to record bacterial growth measurements because an experiment takes hundreds of hours. In a better lab, they would go home at five in the afternoon, and the system would keep running and logging results overnight. That is the kind of change I want to see at scale.

Aayush:

I would also add democratization: at the start of the computing era, mainframes sat in special rooms and only a few people got access. Today almost everyone carries a powerful computer in their pocket. I think something similar could happen in science.

In a world where AI scientists work well, the ability to design simple experiments, run them and interpret the results could become accessible to many more people, including students and hobbyists. A child asking “why does this behave this way?” could run a small experiment and get a serious, structured explanation of what happened and what it implies. That kind of accessibility would not only train future scientists but also raise the general level of curiosity and understanding in society.

After these projects conclude, will ARIA publish what you have learned?

Ant:

We will not publish detailed results for each individual team, but we do intend to publish a high-level synthesis of what we have learned. That will include our sense of the current state of the art, where systems are actually delivering versus just promising, and where we see major opportunities and gaps.

This is not just useful for scientists. If AI scientists become real and have the impact we expect, they will matter to everyone. So we want to make sure that the broader public narrative is informed by reality rather than by hype alone. That means being transparent, at the right level of abstraction, about what works and what does not.

We are also thinking hard about ethics, governance and the role of humans in the loop as these systems evolve. The goal is to help create the wave we want, not to be surprised by the one we get.

Thank you both, excited to see the final report.

If you are interested in keeping up with ARIA’s progress in this space, follow ARIA here.

Did we miss anything? Would you like to contribute to Decoding Science by writing a guest post? Drop us a note here or chat with us on X.

Subscribe now

What we read

AI chips are getting hotter. A microfluidics breakthrough goes straight to the silicon to cool up to three times better [Catherine Bolgar, Microsoft, Sept 2025]

Datacenters, computers, phones, chips, all electrons zipping around to make the world work and it’s no surprise to anyone in modern civilization that it creates heat. Cold plates have been the industry standard in cooling down these heat generating processes in industry to date, but on September 24, 2025, Microsoft unveiled a successful test of a new method that cooled 3x more effectively, using microfluidics!

Microsoft built the microfluidics system into the chip, as shown above, and cooled a server running core services for a simulated Teams meeting. This technology reduced the maximum temperature rise of the silicon inside the GPU (specific to the chip being tested) by 65%.

The coolest part of this innovation is that it was bio-inspired. Clearly the photo shows a complex pattern, and it was actually inspired by veins in leaves or butterfly wings with the help of AI from a Swiss startup called Corintis.

I can personally appreciate the scale of the microfluidics channel because a system at this scale and complexity is no easy feat. The engineering required to ensure no clogging or creating a structural problem in the chip itself when engraving the channels must be acknowledged. Four design interactions took the team a whole year. It is also just the beginning, as microfluidics as a cooling mechanism can bring a coolant much closer to where power is consumed compared to before. In other words, it might just allow for a new age of chip design, 3D chips!

To finish off this cool announcement is a much needed reality check in the fast paced world of software and AI: “Hardware is the foundation of our services,” says Jim Kleewein, a technical fellow at Microsoft 365 Core Management. I couldn’t agree more that at the core of our everyday lives is a physical reality that the digital world is built on, and hope to see more innovations that will help manage the energy and water usage during this AI boom.

Rethinking human roles in AI warfare [Davidovic, Nature Machine Intelligence, Comment, October 2025]

A ‘human in the loop’ has been hailed as the key solution to many of the ethical risks that arise from the use of AI-enabled weapons. However, whether and when we should rely on human engagement for ethical risk mitigation depends on the purpose and type of engagement. The author argues that scholars are not explicit about the purposes behind their calls for meaningful human control, therefore, a taxonomy of the types and purposes of human engagement can clarify these decisions.

The main reasons why academics and policymakers insist on human involvement in AI-enabled weapons include:

Double-checking outputs, especially in decision-support or targeting tools

Responsibility attribution. Only human agents can be held accountable.

Human dignity. Policymakers argue that without human judgement it fails to treat victims as having dignity.

Institutional stability. Institutions depend on human rules and therefore human control is necessary to preserve these institutional structures.

Justice. Morally consequential decisions must be made by a human.

A second input, the author argues, is the five types of engagement: human control, human judgement, human deliberation, preserving intent, and preserving values. This taxonomy is important as the decision, training an organizational structure might support one purpose or engagement (such as safety via human control), but not others.

A third input for ‘human in the loop’ actionable policies is the ‘unit of analysis or the unit of governability’. The targeting process includes six stages: “(1) the commander’s intent, objectives and guidance; (2) target development, validation, nomination and prioritization; (3) capabilities analysis; (4) the commander’s decision, force planning and assignment; (5) mission planning and execution; and (6) combat assessment”. A practical policy should also specify which stage of the targeting process should focus on.

In conclusion, without clarity across these three inputs, policies cannot have institutional and technical compliance for AI-enabled weapons.

Community & other links

Factory Settings, a recently launched IFP newsletter on the success of the CHIPS Act.

Charles Yang wrote a Philanthropic Agenda for Accelerating Autonomous Labs. It describes different avenues and programs of how philanthropy can advance autonomous labs.

What we read

Expert Evaluation of LLM World Models: A High-Tc Superconductivity Case Study [Guo et al, arXiv, November 2025]

As a PhD researcher myself I see the field converge towards relying solely on LLMs for idea generation, review of work, and answers. At the surface responses appear well researched, technical, and generally accurate. But are these answers accurate enough?

To answer how well suited LLMs are at answering scientific questions with PhD-level expertise Guo et al focused on the field of high-temperature superconducting (HTS) materials. Working together with 63 different experts across more than 15 different institutions the authors set out to evaluate how factually correct, well-balanced, and coherent responses were with the field. Four commercially available LLMs (ChatGPT-4o, Perplexity, Claude 3.5, Gemini Advanced Pro 1.5), NotebookLM, and a custom built retrieval-augmented generation (RAG) were compared (Figure 1). At the time of research (end 2024 - early 2025) only Perplexity and the in-house built model were able to return figures.

Figure 1: (a) Distribution of physics-based categories covered by the 67 questions experts set for models to answer, presenting a diverse and well-balanced set of topics. (b) Sample question posed to each LLM model evaluated. (c) Response of model 6, the custom built retrieval-augmented generation (RAG) model that was presented by authors to complement the four commercial closed models, and NotebookLM. (d) Inset of figure 3 presenting mean scores and standard errors of models for five grading criteria defined. The number of individuals grading each aspect were listed in table ‘f’.

What they hoped for the LLM to achieve was an objective understanding of literature that could help researchers entering the field to understand fundamental theories, experimental insights, and how these synergistically interacted with one another.

What they found LLMs were:

1. good at, was answering questions with well-defined quantities (numerical responses required).

2. insufficient at, was:

   1. Understanding concepts across papers. If key words were not explicitly mentioned, concepts were not picked up on. Connections that any researcher would make were not made.

   2. Contextual understanding, across space and time. Papers that contradicted one another were not considered in a wider context; publications that later resolved - or proved prior theories wrong - were not evaluated within latest experimental insights.

   3. Citing accurate sources. For web-interacting models in particular this remained a concern.

   4. Understanding of images and deeper engagement with literature. Even when image captions were embedded into the model often times models still struggled to return the correct image.

So, can we rely on models to provide us with truly expert answers? Probably not yet, even considering this research dates early 2025. The need for models to understand, beyond superficial synthesis of information, remains a limiting bottleneck to acquiring expertise. Inability to engage with images adds to this challenge. Reasoning capabilities, curiosity-driven research, ability to question, and deep expertise continue to be extremely valuable, if not more so in an age where transcription and paraphrasing become the standard norm.

Community & other links

A reading list from Cosmos Institute’s AIxScience Seminar.

alphaXiv raised a $7M seed round. They are building a practical, collaborative platform to interact with scattered AI research, and we are proud to call them our partners.

Subscribe now

Did we miss anything? Would you like to contribute to Decoding Science by writing a guest post? Drop us a note here or chat with us on X.

Subscribe now

What we read

Looking for Hidden Gems in Scientific Literature [Ulkar Aghayeva, Elicit, Oct 2025]

Aghayeva surveys the field of literature-based discovery (LBD), which shows certain promise when it comes “to search for and reveal already existing, but still hidden, links between concepts, findings, questions and answers within scientific literature that would otherwise take much longer to stumble into.” She examines why increasingly sophisticated computational methods have failed to deliver the discovery breakthroughs their technical capabilities might suggest.

She contends that LBD’s core dysfunction rests within “the evaluation problem.” The field has historically been benchmarking against a handful of manually-discovered cases and attempts to bypass this bottleneck have resulted in “co-occurrences [that] are trivial or noisy, like pairings with generic terms. A high score assigned by an LBD method thus doesn’t necessarily point to a novel, valuable and generative connection.” This creates a situation where “it has been easier to make a technological contribution to LBD, by developing a new algorithmic method, than to put together a high-quality annotated dataset.”.

Beyond the evaluation problem, Aghayeva articulates how different types of creativity map onto LLM capabilities in ways that constrain what LBD can achieve. She argues that LLMs excel at combinatorial and exploratory creativity but lack transformative creativity that “alters the concept-space itself.” This leads to the memorization versus generalization trade-off in LLMs, where models can produce interesting combinations from their training data but cannot build entirely new frameworks. The implication is that LBD may be inherently limited to incremental discoveries.

Here she proposes a solution put forth by Gwern: a “daydreaming loop” that continuously samples concept pairs, generates connections, and filters for value. However, as Aghayeva notes this requires accepting the computational cost of producing a large number of pairings that just won’t be interesting. Yet “if we already knew how to predict the value and interestingness of associations, there wouldn’t have been a need to traverse the entire space of possibilities to begin with.” Moreover, the most valuable discoveries might be precisely “the most far-flung and low-prior connections,” so perhaps no real shortcuts exist.

A Sensing Whole Brain Zebrafish Foundation Model for Neuron Dynamics and Behavior [Vegas et al, arXiv, Oct 2025]

The biggest challenge with understanding the brain is being able to visualise neuronal firing processes and simultaneously link these to actions, in real time. Whilst mice remain the key model of choice for neuroscience, larval zebrafish offer the unique advantage of being transparent. A see-through body: engineered for cells to fluoresce upon influx of calcium ions, and acting as a blueprint of firing patterns of the brain.

In the recent preprint shared by Vegas et al, this unique capability of peering through a fish’s body is taken advantage of for recreating, and subsequently learning from, neuron patterns. Following the workflow in figure 1, the authors made use of a larval zebrafish whole-brain calcium-imaging dataset. Rather than training it on raw fluorescence however they train the sparse brain model (SBM) developed on spiking statistics. These are extracted using a poisson transformation, applied after the causal self-attention model of CASCADE is applied.

Figure 1: (A) Data preprocessing and sparse brain model (SBM) architecture. The SBM consists of two layers; the first for spatial activation across neurons, and the second for temporal activity within each neuron. The peripheral neural model allows connecting traces to behaviour; using gradient descent on this allows inferring behaviour from patterns. (B). Comparison of ground-truth experimental firing patterns obtained from the calcium imaging dataset, to model predicted neuronal patterns at timepoints 0-4. The SBM used a small context window of 4 s ( = 12 steps, as recordings are sampled at 3 Hz) per inference, hence why four snapshots were sufficient to encapsulate the entire dataset from which the model is pulling historically for training.

Briefly, let’s explain causal self-attention, as it is a critical factor to making time-inference in a dynamic and real-time way possible. How does it work? Self-attention in a transformer lets each element in an input sequence account for other elements in the same sequence to compute its representation. If the sequence is states in time, a state at time t can thus query states t’. The causal factor of causal self-attention comes from the superposition of a causal mask: an element that prevents nodes (neurons in this case) in the net from considering future states t’. Thus, a node can only access its entire firing history and current state, such that t’ t. The model is called dynamic because it learns from the output of the neuron spatio-temporal layer: any output from this layer will represent how each neuron fires according to other neurons being active. Thus, the temporal layer considers the history of a single neuron firing, but indirectly also accounts for all other neurons firing. This is what makes the proposed SBM uniquely capable of being able to maintain single-neuron interpretability whilst also scaling to the whole net (or whole brain).

Beyond contextualising biological activity across scales, the SBM took interpretation of brain activity a step further by linking output traces to a peripheral neural model (PNM). The PNM allows connecting brain activity traces to behavioural outputs, which in of itself is beneficial to understanding what patterns give rise to which behaviors. Yet once a coherent pattern is understood to induce a behaviour, it can then be used to recreate that behaviour. In this case gradient-based synthesis of the PNM made neural pattern <> behaviour matching possible.

So what becomes possible with a foundation model mimicking - and recreating - neuronal firing patterns in larval zebrafish? From the outside the notion of simulating neurons across scales may appear more as a game rather than a tool i.e. ‘guess the fish swimming direction’. And whilst this virtual larval zebrafish foundation model may not be able to infer drug toxicity (yet), it could in future give insight into neural patterns required for specific behavior, acting as a model to understand neurodegenerative disease effects on pattern loss. A step - or swim - in the right direction, towards catching a wave!

Enzyme specificity prediction using cross attention graph neural networks [Cui et al, Nature, October 2025]

Enzymes, the essential molecular machines of life, are defined by their substrate specificity—the ability to selectively recognize and act upon particular substrates. Although many enzymes can act promiscuously, catalyzing reactions with non-native substrates, a major challenge remains: for millions of known enzymes, substrate specificity is still poorly characterized. This lack of information limits both their practical use and our broader understanding of biocatalytic diversity.

Existing machine learning models for enzyme specificity prediction have achieved only limited success, often being restricted to particular protein families. Most rely on features derived from one-dimensional protein sequences or graph representations, overlooking the inherently 3D nature of substrate binding and the intricate interactions between enzyme and substrate. Current tools such as CLEAN, ProteInfer, and DeepECTransformer struggle to distinguish between enzyme reactivity and substrate specificity within the same EC numbers—an enduring challenge in biocatalysis. Moreover, many previous approaches represented enzymes and substrates as separate embeddings before concatenation, which hindered the accurate capture of their detailed interdependencies.

To address these limitations, researchers at UIUC developed EZSpecificity, a general deep learning framework for predicting enzyme substrate specificity. The model uniquely integrates sequence information, 3D enzyme–substrate complex structures, and the active-site environment. A key innovation is its cross-attention–empowered SE(3)-equivariant graph neural network (GNN) architecture, trained on a comprehensive, custom-built database of enzyme–substrate interactions (ESIbank). Unlike earlier methods that treated all interactions equally, EZSpecificity’s cross-attention layers selectively emphasize critical amino acids and atoms, reducing noise and enhancing generalizability. Its SE(3)-equivariant GNN encoder also captures the atomic microenvironment within the catalytic site.

Experimental validation showed that EZSpecificity substantially outperformed existing models, including ESP, achieving 91.7% accuracy in identifying the single reactive substrate—compared to ESP’s 58.3%. The framework also demonstrated strong generalization across diverse protein families and the ability to predict outcomes for previously unseen enzymes and substrates. Future work will focus on integrating dynamic binding information to further boost predictive power.

EZSpecificity also shows promise for studying biosynthetic gene clusters (BGCs), linking genes to their corresponding intermediates with up to 66.7% accuracy in identifying the correct target enzyme. This represents yet another compelling example of AI’s potential to accelerate drug discovery through large-scale biochemical data analysis.

Community & Deals

Corin Wagen shared his thoughts on AI Scientists. He defines them as “capable of some independent and autonomous scientific exploration” and discusses their presence, opportunities, and limitations.

Fireworks AI raised $250m in Series C funding at a $4b valuation. They’re developing a platform that allows companies to build internal generative AI capabilities.

Substrate emerges from stealth with $100M funding to rival ASML with its novel chipmaking technology.

Field Trip

Subscribe now

Did we miss anything? Would you like to contribute to Decoding Science by writing a guest post? Drop us a note here or chat with us on X.

Subscribe now

What we read

Data-driven de novo design of super-adhesive hydrogels [Liao et al, Nature, August 2025]

AI in science has created a lot of buzz, with numerous companies spinning out especially in the biotech world with better drug design as a goal. Some scientists have been skeptical about the effectiveness, choosing to wait for late stage phase trial results before passing a verdict. However, this paper looks at AI being used in material science, showcasing a successful deployment of AI to help scientists better design and translate the AI’s insights into real world materials.

Liao et al. were able to create an adhesive hydrogel with a maximum adhesive value exceeding 1MPa, a value that is an order of magnitude higher than previous materials. This discovery is relevant for numerous applications, from medical applications to underwater exploration. By cycling through data mining, experimentation, and machine learning approaches, they found an optimal combination of copying nature but better.

A total of 24,707 proteins belonging to 3,822 different organisms were sourced from the National Center for Biotechnology Information Database. This is when the scientists took a page from nature, choosing to focus on the 200 species with the most adhesive proteins since evolution must have given them the better proteins. The 20 basic amino acids were categorized into 6 groups, hydrophobic, nucleophilic, acidic, cationic, amide, and aromatic for ease of modeling and also to simplify the synthetic options. To replicate the proteins, six monomers were chosen to substitute the more complex biological counterparts. From analyzing the properties of the proteins in the 200 species, 180 different hydrogels were experimentally synthesized and tested. These hydrogels were represented by the six monomers and fed to 9 different machine learning models, where Gaussian process (GP) and random forest regression stood out to be the best base models. After more optimization, three super hydrogels were synthesized and tested, the second of which is shown being tested in the figure below!

Magnetically Responsive Microprintable Soft Nanocomposites with Tunable Nanoparticle Loading [Sun et al, arXiv, September 2025]

Previous Decoding Science posts* have mentioned this before - and will mention it again here: material innovations are the biggest bottlenecks to diversifying robot applications across tasks and scales.

In this recent preprint Sun et al. fabricated a microscale magnetically actuated hydrogel. Using an extensive suite of characterisation techniques, the paper brings with it a healthy dose of materials engineering, and insight to learn from. In a field where application and innovation speed often feel like they surpass reflection and diverse problem-solving method exploration, the bottom up approach of redesigning a manufacturing method from scratch was what led to a new composite actuator: flexible, manipulable, and magnetically strong.

In a two-step process Sun et al. first print a poly(ethylene glycol) diacrylate (PEGDA) hydrogel using two-photon polymerisation (2pp) and then chemically form magnetic nanoparticles in situ (figure 1 A-B). Taking advantage of the fabrication technique - which rasters lasers over the hydrogel and crosslinks monomers in doing so - they were able to change crosslinking density in a highly localised manner. Subsequently, dipping the 3D-printed construct into first a bath of iron ions, and then a bath of ammonium hydroxide solvent, allowed iron oxide nanoparticles (IONPs) to co-precipitate in situ. This in situ formation meant NPs did not interfere with lasers during printing; a key improvement over previous proposed magnetic hydrogel microfabrication techniques.

Figure 1: Fabrication and sample application of the microscale magnetically actuated material. A) Diagram of the two-step synthesis process used for embedding iron oxide nanoparticles (IONPs) into a pre-polymerised and 3D printed hydrogel matrix. B) Effect of crosslinking density on diffusion of iron ions and ammonia, influencing co-precipitation efficiency of IONPs in situ. C) Effect of sphere size on magnetic deflection for 2% crosslinked spheres of varying sizes, all attached to 18% crosslinked ‘arms’ and a 51% crosslinked base. ii. and iii. visualise pillars without a magnetic field, and in the presence of a fixed permanent magnet. D) Effect of crosslinking density on magnetic deflection for spheres of fixed size. Base dimensions and magnetic field setup (ii. and iii.) were kept constant to those of ‘C’. E) Asymmetric microscale gripper fabricated with two 2%, and two 51% crosslinked spheres. F) Application of a payload and magnetic field demonstrating gripper function.

The possibility to influence hydrogel crosslinking density by modulating laser power during printing also affected ion and solvent diffusion. In effect, a higher crosslinking density yielded lower magnetic strength, as IONP co-precipitation could not penetrate the dense matrix of small pores, thus remaining at the surface. Understanding these relationships was essential to engineering - in a highly controllable manner - soft robotic ‘gripper’ arms. As shown in figure 1 C-D, IONP-infused spheres were optimised for size and magnetic strength. In taking advantage of the ability to control magnetic response through crosslinking, two different sphere ‘arms’ were made. This asymmetric design gave the gripper both a soft ‘guiding’ force from the 2% spheres (low crosslinked <> high magnetic response), and stronger ‘support’ force from 51% spheres (high crosslinking density <> medium/low magnetic response).

Projecting into the future the authors envision applications in the soft robotics and medical space. In parallel, I propose imagining what an assembly line of magnetic actuators could look like. The authors demonstrated that two different components, printed with the same crosslinking density, could respond differently provided a magnetic field was locally applied. Taking this further, one could imagine a mini-factory, where each ‘gripper’ is a module along the assembly line: sorting stem cell organoids, applying mechanical force to activate biological pathways, and sorting only the mature and correctly differentiated organoids once more. The opportunity space of soft grippers in on-benchtop assembly systems is vast; provided outstanding challenges in biocompatibility, composite stability, and controlled force calibration can be shown these could introduce entirely new ways of working with biology, using tools that complement the properties of systems they are creating.

*See DS 003 on PINNs for programmable origami metamaterials, and DS 001 on ultrasound-driven autonomous robots.

Physics-Informed Neural Networks for Real-Time Deformation-Aware AR Surgical Tracking [Harper et al., biorXiv, September 2025]

Surgical navigation driven by augmented reality is unreliable when using conventional methods due to soft tissue deformation, which can lead to significant, unacceptable errors in surgical scenes. Similarly, data-driven models have led to anatomically implausible results when training data are limited or out of distribution and do not cover extreme deformations.

Physics-Informed Neural Networks (PINNs), which combine physical laws with neural models, have not been readily applied to surgical AR tracking. The authors introduce a PINN that embeds Finite Element Modeling (FEM) elasticity constraints to achieve real-time, deformation-aware AR surgical tracking. FEM splits a complex object into many small, discrete objects, solves the partial differential equations on each, and combines the results to approximate the overall object’s real behavior.

The PINN+FEM combination led to a lower target registration error (error between virtual model and real patient) of 1.1mm compared to FEM-only’s 1.8mm and ICP’s 2.9mm.

Community & Notable Deals:

The British Research Agency ARIA opened calls for projects about AI scientist systems, offering up to £500,000 for each project.

Lila Sciences raised $350M in Series A to build a scientific superintelligence for autonomous research labs, the so-called “AI science factories.” Also covered on Decoding Bio.

General Intuition emerged from stealth with a $133M seed round, building a foundation model with improved temporal and spatial reasoning by training on video games.

Subscribe now

Did we miss anything? Would you like to contribute to Decoding Science by writing a guest post? Drop us a note here or chat with us on Twitter: @pablolubroth @ameekapadia

Subscribe now

What we read

Thoughts on The Curve [Nathan Lambert, Interconnects, October 2025]

Lambert puts together some notes following a conversation at The Curve, an AI conference hosted earlier this month by the Golden Gate Institute. The first half of the article is of particular interest here because it attempts to justify certain timelines on AI for Science.

Here, I specifically want to focus on his argument for how the “AI Research Engineer” role resists clean automation timelines. Per Lambert, automated RE’s (agents that can see through the whole sequence of the research methodology) are feasible within the next few years with some caveats. He argues that RE is a broad category and its components will fragment at variable rates. Some ideas will be automated quicker, especially with more computing power, while others will see a narrowing but present bottleneck where human insight is still necessary – but “to check the box of automation, the entire role needs to be replaced.”

Beyond technical capabilities, Lambert also identifies how different modes of scientific discovery resist automation asymmetrically. Cross-pollinating existing fields would be much easier for agents than having truly transformative breakthroughs: “being immersed in the state of the art and having a brilliant insight that makes anywhere from a ripple causing small performance gain to a tsunami reshaping the field.” Moreover, the marketplace of ideas where researchers convince colleagues to pursue specific directions will maintain its human core even as everyone gains “superpowers on making evidence to support their claims.”

Lambert’s final point on the friction inherent with autonomous AI REs is based on the “inevitable curse of complexity” where each advance in AI systems needs more and more elaborate infrastructure, tooling, and wrappers. He is of the opinion that this will create a sort of drag that ensures progress “will feel much more linear rather than exponential” going forward.

What we read

Active-Learning Inspired Ab Initio Theory-Experiment Loop Approach for Management of Material Defects: Application to Superconducting Qubits [Chaudhari et al., arXiv, October 2025]

Superconducting quantum interference devices (SQUIDs) were first introduced in 1985 with the discovery that macroscopic tunneling effects could occur between two layers of superconductive material separated by a thin insulating barrier [1]. As John Clarke, Michael Devoret, and John Martinis demonstrated, electron pairs were able to move coherently - and without resistance - through the material lattice: this is the foundation of superconductivity. Yet what differentiated this sandwiched structure from other materials was its assembly into a quantum device. In creating the first Josephson junctions (JJs) setup, they allowed electron pairs to tunnel through the insulating layer - hopping from one superconducting material layer to the other - given sufficient current was applied. Being able to control electrons in such confined environments was the first step to translating theoretical principles of quantum computing to the real world.

Since then, innovations in cryogenic control, material layering combinations, and algorithmic organisation have incrementally moved the field forward - albeit at cryogenic speeds, one might note. But why has progress been slow: why do we still carry around silicon chips as opposed to qubit-based devices in our pockets and lives? Because the biggest bottleneck remains materials selection. Overcoming the challenges of decoherence and noise is critical to enabling the stable differentiation between states - effectively superimposing a computable state.

To this end, the recent active-learning framework proposed by Chaudhari et al. provides a unique approach to find new material combinations. In the paper, they combine density functional theorem (DFT) computational predictions and a trained logistic regression model with empirical discovery. This ‘Material Genome Loop’-inspired setup was used to identify new materials to interface with niobium. Given the use of a logistic regression model, the input was equated to a set of continuous-value descriptors, and the output was computed as a probability score rounded to a binary predicted value. Descriptors used were broken down into 1. defect formation energies calculated with DFT, and 2. thermodynamic descriptors (fitting parameters). Thermodynamic descriptors, in particular, were computationally shown to be the dominant effect in determing whether an interface would oxidise or not. Thinking of an oxygen atom as being in a simulated game, the different ‘moves’ an oxygen atom could make were: i. moving from the candidate metal into (a) the niobium layer, (b) a niobium oxide cluster, or ii. moving from a metal oxide into (a) an interstitial site in the lattice. The latter effectively freed the oxygen atom to migrate through grain boundaries in the material and oxidise the interface.

Yet what strengthened the paper setup and results were the tight integration of theoretical predictions with experimentally validated results. As shown in figure 1, predicted probabilities of forming an oxide (inner circle color) were compared to experimentally observed oxide formation (outer circle color). The active-learning framework of taking model-recommended candidate materials → testing these experimentally → using results to further train the model allowed narrowing the search scope and identifying top candidates. Combined with materials expertise of selecting a crystal structure to match that of niobium, the authors were able to identify top Josephson Junction material candidates to prevent oxide formation.

Figure 1: Energy to move into a metal interstitial compared to oxygen vacancy site. (a and b) Green regions indicated that an oxygen migration pathway was inhibited. Overlap in regions indicated more than one region was inhibited; the darker the green tint the more pathways were inhibited. From this coloring scheme, the top left-hand corner emerges as most promising to prevent oxide formation. Red and green circles outline reflected if Nb oxide formation was suppressed (green) or not (red) experimentally. The inner circle color plotted the probability of forming an oxide as predicted by the logistic regression model. (c) Lattice mismatch % between each candidate material and the BCC niobium crystal structure, plotted against oxide formation energy.

So why does it matter that zirconium (Zr) was ranked best, followed by tantalum (Ta), hafnium (Hf), and scandium (Sc)? Because to date, most superconductive qubit devices continue using aluminium (Al) as the default material for the construction of Josephson Junctions. However, as demonstrated by Chaudhari et al. using this active-learning framework, aluminium does not emerge as the best candidate to match niobium qubits. In light of lattice mismatch and predicted oxide formation energy per oxygen atom, rational design strategies could be used to re-evaluate material choices in SQUIDs and uncover new superconductor/insulator combinations that move the field forward, faster, one tunneling step at a time.

Discovering new solutions to century-old problems in fluid dynamics [Wang et al., DeepMind, September 2025]

A major unresolved problem in physics is whether certain fluid dynamics equations (e.g. 3D Euler and Navier-Stokes equations) can develop singularities; these are points in time and space where some quantity, like velocity or pressure, becomes infinite. These help mathematicians identify limitations to these equations and help us improve our understanding of fluid dynamics. In fact, the problem of finding a singularity in the Navier-Stokes equations is so notorious that it is one of the six Millennium Prize Problems, which are still unresolved. Unstable singularities (those that any slight perturbation might destroy them) are likely the ones that exist in these two equations.

The team at DeepMind used a physics-informed neural network and a framework that pushes the PINN to near-machine precision to discover a new family of unstable singularities in several fluid equations. They also observed that the speed of the blow up (the process of the property tending to infinity) and how unstable the solution is, are correlated in two of the equations.

Nobel 2025 Laureates: Current Activities in AI - HC

This year’s Nobel announcements honored foundational science rather than AI methods. In Medicine, Mary E. Brunkow, Fred Ramsdell, and Shimon Sakaguchi were recognized for discoveries establishing peripheral immune tolerance and regulatory T cells (FOXP3). In Physics, John Clarke, Michel H. Devoret, and John M. Martinis were cited for experiments showing macroscopic quantum behavior in superconducting circuits—work that underpins today’s quantum technologies. In Chemistry, Susumu Kitagawa, Richard Robson, and Omar M. Yaghi were honored “for the development of metal-organic frameworks (MOFs),” porous crystalline materials whose tunable architectures enable applications from gas storage and separations to catalysis.

Sakaguchi’s recent work intersects with AI/ML mainly via single-cell immunology. A 2024 CD4⁺ T-cell atlas uses high-dimensional single-cell data and presents a machine-learning framework to predict autoimmune states from T-cell profiles (Cell Genomics 2024). While the Nobel prize itself is for foundational immunology, this line illustrates how his community now leverages ML for disease stratification and mechanism discovery.

Devoret and Martinis connect to AI through quantum error correction (QEC), where learning-based control/decoding is increasingly standard. Devoret’s group published a 2025 Nature paper demonstrating qudit QEC beyond break-even, explicitly optimized with a reinforcement-learning agent - a direct AI application to quantum hardware. Martinis’ former Google Quantum AI program advanced neural-network/transformer decoders for the surface code on Sycamore-class processors, showing state-of-the-art learned decoding on experimental data (Nature 2024). Together, these works exemplify AI as an enabling tool for quantum control and scalability, even though the Physics Nobel recognizes earlier foundational superconducting-circuit breakthroughs.

Omar M. Yaghi is actively weaving AI into reticular chemistry: his team co-led MOFGen, an “agentic AI” pipeline that generated hundreds of thousands of candidate MOFs and guided the successful synthesis of several “AI-dreamt” frameworks, demonstrating end-to-end AI-assisted materials discovery (arXiv 2025). His group has also advocated for—and helped launch—infrastructure to mainstream these tools, notably Berkeley’s Bakar Institute of Digital Materials for the Planet (BIDMaP), where recent pieces outline how generative AI can streamline lab workflows in reticular chemistry. Earlier group work used ChatGPT for literature mining of MOF synthesis conditions and for building predictive synthesis models, foreshadowing today’s agentic systems (JACS 2023).

Notable Deals:

Periodic Labs emerges out of stealth with a $300 million seed round led by a16z to build an AI scientist that accelerates discoveries in physics, chemistry and other fields.

Axiom Math, founded by a former Meta researcher, has raised $64 million in its seed round to build an AI that can generate and solve complex mathematical problems.

Subscribe now

Did we miss anything? Would you like to contribute to Decoding Science by writing a guest post? Drop us a note here or chat with us on Twitter: @pablolubroth @ameekapadia

Subscribe now

Good morning. We’re playing around with the structure of Decoding Sciences’ newsletter. Let us know what you think!

What we read

PRISM: Capturing the Invisible Art of Scientific Practice [Henry Lee, Off Media, July 2025] - HJ

Scientific protocols are often devoid of important tacit knowledge because papers describe what researchers did, but rarely capture how they did it with sufficient granularity. Things such as the precise pipetting angle or the specific visual cues that indicate culture contamination remain invisible: “scientists know that papers shaped by publication pressures often portray successful outcomes, rather than the countless micro-decisions and adaptations that made success possible.” This can then cause unforeseen delays for researchers who are following the same protocol as they work to rediscover methodological nuances.

To circumvent this problem, Cultivarium has developed PRISM – an AI-powered lab assistant that transforms static protocols into “a multimodal record of scientific practice.” Researchers wear glasses that record audio and video data while they are conducting an experiment. This raw feed is then processed via the model into a step-by-step process that can be read or watched for accurate reproducibility.

Beyond the immediate documentation benefits, this tool also generates training data that could be used to build truly autonomous lab agents: “AI systems that not only execute predefined protocols, but can also adapt to the unexpected, recognize when something isn’t working, and even generate novel approaches to experimental challenges.” [See a demo here]

AI for Scientific Discovery is a Social Problem [Channing et al., arXiv, September 2025] – PL

Channing and Ghosh from HuggingFace argue that the democratization of AI for science requires treating it as a collective social project. The paper argues that there are four critical barriers to its democratization and four potential solutions. This is not an exhaustive summary of the paper, and for those interested, we encourage a full read.

Barriers:

1. Community dysfunction that undermines collaboration:

   1. AI Scientists could become useful co-pilots, but are a counterproductive direction to pursue and to scientific progress as it devalues expertise and contributions of human scientists whose creativity and knowledge remain essential. It also oversimplifies the complexity of scientific practice which is not only about predictive accuracy but also on careful validation, contextualization and theoretical integration, and finally, also undermines science’s purpose: the cultivation of human understanding, not only providing solutions.

   2. Collaboration failures that arise in differences in priorities: for instance, domain scientists care about mechanistic understanding and experimental validation, while ML researchers focus on predictive performance and computational efficiency.

2. Misaligned research priorities targeting narrow applications over upstream computational bottlenecks:

   1. Publication pressure and grant cycles create incentives in the scientific community to fragment and solve domain-specific problems, rather than collectively mobilize around computational bottlenecks.

3. Data fragmentation due to incompatible standards

   1. Hoarding data in proprietary formats, lack of incentives for researchers to focus on data curation and harmonization, result in datasets locked in incompatible silos.

   2. The multimodal, spatial and temporal relationships in scientific datasets resist straightforward tokenization approaches, resulting in poor predictive performance despite the scale of the underlying data.

4. Infrastructure inequities concentrating power within privileged institutions

   1. Academic researchers face a lack of technical infrastructure, which is even more pronounced when compared with institutions across the world.

Solutions:

1. Strengthening collaboration and education across communities via standardized interfaces and APIs, community-driven development and training of domain specialists into ML specialists, and vice versa.

2. Structuring upstream challenges, such as the Vesuvius Challenge & DREAM Challenges, and shared benchmarks.

3. Standardizing and curating scientific data for broad reuse such as the standardization via file formats like CSV lower barriers to collaboration, but also through the development of architectures for scientific data (such as graph neural networks molecular and materials applications by modeling spatial relationships or physics-informed neural networks which incorporate domain knowledge into model design).

4. Building accessible and sustainable infrastructure through the sharing of, not just trained models, but entire scientific AI pipelines, as well as building community-owned infrastructure and sustainable funding.

Community & other links

AI models are using material from retracted scientific papers [Ananya, MIT Technology Review, September 2025]

What we read

The Open DAC 2025 Dataset for Sorbent Discovery in Direct Air Capture [Sriram et al., arXiv, September 2025] - IVdB

When thinking about direct air capture (DAC) we like to think solutions will fall out of the sky to vacuum the vast quantities of pollutants humanity exhausts on a daily basis. In the case of the newly updated Open direct air capture dataset (ODAC25) solutions might not fall out of the sky, but they could appear out of thin air.

Carbon dioxide accounts for ~76% of all greenhouse gases emitted [1]. Rather than dreaming about vacuum-cleaning solutions, the most realistic approach is to start at the source: the CO2 emitting factories. This work, released from Sriram et al., can have an important impact in providing solutions - in the form of rationally selected complex metal organic framework (MOF) materials. Metal organic frameworks are crystalline porous materials linking metal ions to organic ligands. Their high surface area networks with tunable pore size and chemistry make them ideal for active coatings and high-energy reactions.

Building on the previous ODAC23 released version [2], Sriram et al. expanded the dataset to: 1. introduce two new species of absorbents, 2. simulate materials in more realistic conditions, and 3. add chemical diversity to MOFs. First, along with CO2 and H2O, the addition of nitrogen and oxygen allowed accounting for surface oxidation. As is the case with copper roofs turning green from oxidation, so is accounting for MOF spin- and oxidation states relevant to prevent material corrosion. Introducing oxygen into simulations allowed screening of previous adsorbent-material interactions that would have not been accounted for in ODAC23.

Second, in accounting for oxidation states MOFs can be selected for highly specific applications; any that may suffer from spurious oxygen dissociation - where redox reactions can occur - will be flagged for further atomistic or experimental inspection. Performing high-energy Grand Canonical Monte Carlo (GCMC) simulations also allowed accounting for non-relaxed states of adsorbed molecules. These higher energy states are relevant for predicting an adsorption isotherm: a curve that allows seeing how much gas each MOF can adsorb at different pressures. Together, introducing oxidation states and making isotherm sampling possible has ensured that more realistic MOFs are put forth for experimental testing, and higher specificity may arise for environmental conditions they are exposed to.

Figure 1: a) Representation of two different functionalizations of linker amination (bottom), and OMS diamine functionalization (top) introduced into the dataset of metal organic frameworks. IRMOF-74-III was used as a sample MOF. i. Demonstration of ion-paired cooperative formation for a carbamate chain provided CO₂ is adsorbed onto metal amine lattices. ii. Carbamic acid, a sample variation of the reaction in ‘i.’ where instead of a 2:1 stoichiometry only a 1:1 stoichiometry exists. iii. Reacted product ‘ii’ for carbamic acid on MOF. b) Difference in most favourable adsorption energies between functionalized and non-functionalized MOFs for CO₂ and H₂O. c) Energy difference in amine-functionalized, and OMS functionalized, MOFs.

Finally, by introducing two different chemical functionalizations using i. amines and ii. OMS diamine a more diverse material space could be explored. Experimentally both linker- and OMS functionalization have been shown to enhance CO2 adsorption: the latter worked particularly well for environments with low CO2 partial pressures. As shown in figure b) and c) functionalization was found to increase adsorption probability for both CO2 and H2O.

So what are the next steps? Testing out the dataset in various environmental conditions. In the paper ODAC25 was compared against two other pretrained machine learning interatomic potentials (MLIPs). However, it would be extremely interesting to see if predicted MOFs could be validated to indeed have higher absorption efficacy in real-world environments, installed above different industrial plant exhausts. A model could then be trained on the chemical ‘feasibility’ of synthesis, using this as a score to fine-tune and explore the space of MOFs generated in ODAC25 for specific applications. And, if proven correct, MOFs could be designed rationally for different exhaust combinations, providing optimal reaction surfaces that maximize CO2 capture for each plant. Linking MOFs and captured carbon varieties to downstream processes could even result in financial benefits from products that emerge. We stay on the lookout for individuals exploring the space.

Chemists Can Discover New Materials More Quickly With AI [Kirsten Heuring, Carnegie Mellon University News, September 2025] & Design of Tough 3D Printable Elastomers with Human-in-the-Loop Reinforcement Learning [Rapp et al., Angewandte Chemie, July 2025]- HW

Researchers at Carnegie Mellon University and collaborators developed an AI system that accelerates the discovery of new polymers by helping chemists navigate the tradeoff between strength and flexibility—two properties that are traditionally difficult to optimize together. The machine learning model was trained on large datasets of known materials and can propose promising molecular structures that balance these competing attributes. This approach significantly reduces the trial-and-error cycle in the lab, enabling scientists to focus their experiments on the most likely candidates rather than testing thousands of possibilities.

Early applications have already identified a novel polymer with an unusually strong yet flexible profile, which could have an impact in areas like aerospace, biomedical devices, and sustainable packaging. Beyond this single discovery, the researchers highlight that the workflow is generalizable: AI can guide experimentalists toward efficient design strategies, making materials innovation more predictable and systematic. This integration of domain knowledge with data-driven models represents a shift toward collaborative “human-AI chemistry,” where computational tools complement scientific intuition to push the boundaries of material science.

Field Trip

Subscribe now

Did we miss anything? Would you like to contribute to Decoding Science by writing a guest post? Drop us a note here or chat with us on Twitter: @pablolubroth @ameekapadia

Subscribe now

What we read

Blogs

Using AI to perceive the universe in greater depth [Tracey and Buchli, Google DeepMind, Sept 2025] – PL

DeepMind, in collaboration with LIGO (Laser Interferometer Gravitational-Wave Observatory), developed Deep Loop Shaping, a new AI method that reduces noise and improves control in a gravitational wave observatory’s feedback system by 30 to 100x.

Gravitational waves are ripples in the fabric of space-time. These waves are generated by events like neutron star and black hole collisions. LIGO first detected the waves in 2015, verifying Einstein’s general theory of relativity. Since then, the observatory and its astronomers have proven the existence of binary black hole systems and studied the creation of heavy elements like gold.

LIGO measures the origins and properties of gravitational waves, however, the slightest vibration can disrupt its measurements. As the blog states “even from waves crashing 100 miles away on the Gulf coast [can disrupt it]”. To keep the systems aligned, there are thousands of control systems which adapt to disturbances with continuous feedback. Deep Loop Shaping improves the stability of its interferometer mirrors, helping astronomers gather data in greater detail, to understand galaxy evolution.

Deep Loop Shaping uses reinforcement learning and was trained to avoid amplifying noise in the observation band for measuring gravitational waves, and over time, the controllers learn to stabilize the mirrors without adding extra noise. In the future, the same model could be applied to other engineering problems such as “vibration suppression, noise cancellation and highly dynamic or unstable systems important in aerospace, robotics, and structural engineering.”

AI enters the grant game, picking winners [Siddhant Pudeskar, Science, Aug 2025] / When AI rejects your grant proposal [David Adam, Nature, Sept 2025] - HJ

The metascience debate tends to focus on ways that agents will change science from within, on how they will impact the experience of doing science – things like hypothesis generation and publishing. These are the places where AI usage is perhaps the most noticeable and salient. It is perhaps less clear how AI impacts the external scaffolding of research, such as its influence on grant making.

For example, at Imperial College London, the Climate Solutions Catalyst program is using AI as “a big filter” to identify promising research with commercial potential. They trained ChatGPT on a mix of green chemistry studies where some had industry-potential and others didn't. The model was then used to identify 160 commercially viable papers from 10,000 abstracts, a selection that was further narrowed down by human input before distributing no-strings-attached grants to three researchers.

In contrast, the La Caixa Foundation used a model trained on successful applications from previous years to pre-screen biomedical applications and reject those with low success probability before human review. The system flagged 122 of 714 applications as unlikely to succeed; reviewers ‘rescued’ 46 of them, but only 34 of the remaining 638 applications ultimately received funding.

These are far from perfect mechanisms, and there is resistance to their adoption, but it’s interesting to consider the new incentives that AI-assisted grant-making might create for scientists.

OpenAI for Science [Kevin Weil, X, 2025] - HW

Kevin Weil - formerly OpenAI’s Chief Product Officer and now leading this new effort - introduced OpenAI for Science as an internal initiative at OpenAI. The goal? To develop the “next great scientific instrument”: a powerful, AI-driven platform to accelerate scientific discovery.

Though details remain sparse, the emphasis seems clear: OpenAI is shifting toward building tools that empower scientific research. This likely involves integrating advanced AI capabilities into workflows across disciplines—such as hypothesis generation, literature mining, experimental design, and data interpretation—positioning AI as a collaborative co-scientist.

If successful, this initiative could drastically speed up scientific workflows and broaden the scope of what’s discoverable. Framing AI as an "instrument" underscores OpenAI’s ambition to move beyond software tools toward enabling transformative research infrastructure. It’s a bold vision, and we’ll need to watch for follow-up updates to learn how it takes shape in practice.

Share

Papers

Improving atomic force microscopy structure discovery via style-translation [Huang et al., arXiv, Sept. 2025] - IVdB

If one has ever done atomic force microscopy (AFM) the developments in this improved particle probe model (PPM) proposed by Huang et al. are nothing but exciting. To contextualize: a current AFM experiment may last 1-2 hours, but equipment setup, camera alignment, and placing of a nanometer sized silicon-based tip no larger than half a ricegrain can take another 1-2 hours. But what if experiments were no longer necessary? And how close are generative models getting to simulating data that is representative of physical results?

In looking to improve current particle probe models used to predict molecular structure from experimental AFM images, Huang et al. observed two major challenges preventing experimentalists from transitioning into a predominantly bits-oriented world (let’s leave those pipettes and tweezers behind!). First, predicting molecular structure from AFM images remains challenging - even for experts. Often additional calculations such as density functional theorem (DFT) will be required to validate structures observed. However, even then atomistic detail may lack from experimental images, and alignment of results takes time. And second, as with any space exploring the use of generative ML models to create synthetic data, a lack of training data hinders development.

Overcoming these challenges was done in two ways. First, Huang et al. proposed introducing a ‘style translation’ for generated AFM images. Going from a molecular ball-and-stick model to a simulated AFM image is already possible with PPMs. However, these lack the noise, artifacts, and small distortions that can skew predictions; a resulting overly smooth space (Figure d and f) can misrepresent which molecules are present and result in inaccurate conclusions. To this end a modified generative adversarial network (GAN) called CycleGAN was adapted. Unlike how GANs produce images from random noise vectors, a CycleGAN network is limited to go from image-to-image. In this instance the CycleGAN was set up to go from simulated AFM image → to → experimental AFM image. In applying a styled transformation - rather than a manually introducing artifacts - such as ‘salt and pepper noise’, background gradients, or cutting of random parts of the image - the distribution of data in the image now more closely resembles real experimental data, as the artifacts come from a generated transformation themselves.

To confirm whether generated images indeed more closely resembled experimental results - and were atomically accurate - a second insight enabled surpassing training data constraints. As the dataset used in the CycleGAN did not have the same simulated AFM results as were in the experimental AFM dataset, atomic labels could not be given to simulated molecular graphs. However, Huang et al. noted that even if labels were not available, structures should still be physically meaningful, and as such consistent with DFT calculations. Thus, by comparing the statistical distributions of layer 1 and 2 in the simulated AFM water dataset they were able to calculate local structure distributions and infer from these structural differences (Figure g-i).

Where does this leave the work? At present Huang et al. only demonstrated this for water molecules; a molecular structure that has been explored since ~2015 in AFM. Whether the technique can translate into resonance-based imaging of more complex molecules with equally high resolution fidelity is to be seen. However, what a realistic AFM image generator would unlock is clear: a new way to produce possible molecular structures, ideal for - in the future - exploring dynamic mechanisms of actions in physiologically relevant environments.

Figure 1: Diagrams, simulations, and experimental data of water molecules deposited on a (111) plane of gold. a) Diagram demonstrating the AFM field of view expected from the bilayer of water formed. b) Side view of ‘a’. c) Density distribution of molecules along the plane. It is expected a higher density of water molecules will be found near the gold surface where it is energetically more favourable to be. d) 3D simulated image generated from experimental data (m) recorded. e) Particle probe model (PPM) 3D simulated AFM image to which a style transformation G_u(u) was applied prior to restacking and constructing the image. f) Discrepancies between conventional PPM simulated results predicting a bilayer distribution of water molecules (dotted line), and style transformed and simulated results (smooth line), showing more noise is accounted for in the style transformed model. g-h) Predicted distributions of water molecules in the first (red) and second (blue) layer of the water bilayer on the surface of the gold substrate. In g) is the distribution of free OH bonds perpendicular to the surface, h) the distance between donor-acceptors and hydrogen bond angle, and i) the probability of a tetrahedral shape existing for water molecules present.

Notable deals

Radical Numerics has launched out of stealth. The company is building an engine for recursive self‑improvement: AI that designs and refines AI, accelerating discovery across science and industry.

What we liked on socials channels

Subscribe now

Did we miss anything? Would you like to contribute to Decoding Science by writing a guest post? Drop us a note here or chat with us on Twitter: @pablolubroth @ameekapadia

Subscribe now

What we read

Blogs

NSF and NVIDIA partnership enables Ai2 to develop fully open AI models to fuel U.S. scientific innovation [NSF, Aug 2025] – MB

Last Thursday, the National Science Foundation and NVIDIA announced that they are investing $152 million ($75 and $77 million, respectively) for the Open Multimodal AI Infrastructure to Accelerate Science (OMAI) project led by the Allen Institute for AI (AI2). As part of this project, AI2 will develop open multimodal models trained on scientific data and literature with the dual goals of 1) releasing models that support scientific discovery and 2) creating open infrastructure to support AI research. From the press release, the aim is to develop models that "will enable America's researchers and developers to process and analyze research faster, generate code and visualizations, and connect new insights to past discoveries, accelerating breakthroughs across materials science, biology, energy and more". AI2 has previously developed notable open models, such as the language model Olmo 2 and multimodal model Molmo, with full training details, data, code, and model weights released openly.

Publishing for machines: a necessary revolution [Andrew White, diffuse.one, Aug 2025] – HJ

One consistent frustration that has come up with the introduction of “deep search” in LLM chat interferences is that agents can’t access paywalled scientific research articles. This limits their scope in some ways: if “I make scientific agents that do keyword searches and download papers…[then] just like normal researchers, these agents hit paywalls and have trouble downloading articles.”

Even if everything were open access, White argues that current papers are terribly structured for machine consumption. PDFs scatter text, figures, and tables with no standardized metadata, and the actual scientific process – including failed experiments and iterative refinements – gets quite sanitized by the time it reaches the final article. This broad issue of metascience will get worse with the influx of AI-generated research – without public infrastructure designed for machines who become readers and writers of scientific literature.

White then contends with the question of what it would look like if we were to make “papers for machines” (without removing the “human narrative”). He has five main proposals: moving research quickly to publication, outlining the main hypothesis at the very beginning (making clear the questions one is asking), allowing access to a minimally processed but intuitive “raw-ish” version of the data, harmonizing the relationship between the results and their supporting evidence, discussing how the study impacted the author’s opinion. He also talks about including an automated peer review that looks at the “rigor of science,” while debates on the merits of the science itself take on a new, perhaps more public form.

Using X-Labs to Unleash AI-Driven Scientific Breakthroughs [Caleb Watney, IFP, Aug 2025] – PL

The current US research funding mechanisms are ill suited for the infrastructure-heavy, interdisciplinary nature of AI-driven science. This is mainly due to the PI-led, project-based grant structure of funding, institutions that are hindered by bureaucracy, rigid scopes and lack of long-term vision. AI-driven research requires broad expertise and shared infrastructure, which project-based finance does not allow. Some of the most important AI for Science developments were funded through novel funding models: “Evo 2 was developed at the philanthropically funded Arc Institute. AlphaFold came out of DeepMind. GNoME was built at Google.” These projects succeeded because they combined “deep infrastructure, interdisciplinary teams, and stable, long-term support”.

The authors propose establishing 25 independent X-Labs that would operate like FROs supported by grants of $10-50M per year over seven-year cycles. No more than 70% of the labs would renew into a second term to ensure continual institutional experimentation. Awards should be open to applicants outside academia.

Share

Papers

Physics-Informed Neural Networks for Programmable Origami Metamaterials with Controlled Deployment [Sukheon Kang, arXiv, Aug 2025] – IVdB

In the current field of robotics, the prevailing paradigm remains that of human mimicry; designs are dominated by anthropomorphism. When venturing outside of anthropomorphized robots for complex tasks however training data remains the biggest hurdle in designing new shapes with desired function. Yet what if training data was not the constraining factor?

This is what Kang et al. set out to explore using a physics-informed neural net (PINN). In making use of a PINN this work presented two approaches to designing the elegant cylindrical folding pattern that made up a conical Kresling origami (CKO) shape. These were a forward prediction approach and inverse design approach, both of which used a fully connected feedforward neural net (NN) model. As shown in the inset D in the figure below these consisted of 2 hidden layers and 128 tanh-activated neurons.

By encoding equations solving for mechanical equilibria directly into the training process the need for training data dissolved, as only physically possible solutions emerge. In the forward prediction model five geometric parameters were constrained: the top (a) and bottom (b) edge lengths, the mountain (c) and valley (d) crease length, and the internal angle (β) between the bottom edge and adjacent mountain crease (figure A-B). By keeping these fixed only the input height h could influence resultant CKO shape, as height had a direct correlation to angle (φ) of torque. Like a seashell, the more spiraled the structure the smaller the height - as the absolute area of the 2D paper remained constant. In this manner forward prediction gave insight on what a given geometry could achieve.

In contrast, in the inverse design NN behavior could be specified, with a designed geometry emerging as output. By also allowing the number of unit cells n to vary, function-level specifications and multistability were directly programmed; equilibrium constraints defining torque-free equilibrium (L_phys), energy curve matching (L_target), geometric feasibility (L_penalty), and variational consistency (L_EL) ensured designs remained physically possible. Implementing inverse design thus allowed starting with a proposed shape and solving for the number of folds required by CKO principles to achieve the desired result.

So what does this mean for the future of designable materials? That the next generation of aerospace systems, and soft robotics could gain memory-like behaviors, and undergo shape changes not seen before due to mere constraints in CKO design.

What we liked on socials channels

Field Trip

Subscribe now

Did we miss anything? Would you like to contribute to Decoding Science by writing a guest post? Drop us a note here or chat with us on Twitter: @pablolubroth @ameekapadia

Subscribe now

What we read

Blogs

Aeneas transforms how historians connect the past [Google DeepMind, July 2025] – HJ

A new study published in Nature introduces Aeneas, an AI model specifically designed to contextualize ancient Roman inscriptions. Trained on a dataset of over 176,000 Latin inscriptions, the multimodal neural network can restore gaps in texts with 73%, can geographically attribute input with 72% accuracy in geographical attribution, and can provide dates for a given text within a 13 year range of expert estimates.

To achieve these results, Aeneas uses “embeddings”: a technique that converts each inscription’s textual and contextual information into a numerical representation or “fingerprint.” This encoding captures linguistic patterns, temporal markers, geographical indicators, and relationships to other texts, allowing for rapid comparison with and retrieval of similar inscriptions from the database.

“When Aeneas examined the inscription on a votive altar from Mainz, Germany, it correctly estimated its date to around 214 AD, identified it as produced in the Roman province of Germania Superior, and proposed restorations for its damaged text. Sommerschield describes these results as “jaw-dropping moments” for the team.” [The Art Newspaper]

These are exciting findings especially since traditional epigraphic analysis can take researchers weeks or months. Aeneas highlights contextual parallels much faster which massively increases the surface area of materials available to scholars in the humanities.

OpenAI’s open sourced models - gpt-oss [OpenAI, August 2025] – HW

OpenAI has finally released fully open-source checkpoints that perform on par with the proprietary o4-mini line. The larger variants grab the headlines, but the lean gpt-oss-20 B (20.9 B parameters, 3.6 B active) is the real workhorse: a 12.8 GiB file that fits on a 16 GB gaming GPU, retains the 128 k-token context window, and lets users toggle among low, medium, and high reasoning modes. In its high-reasoning setting, it reaches 85% on MMLU and 74% on GPQA with tool use: well ahead of similarly sized open models (Qwen 2.5-VL-32B: 78% / 46%; Mistral-Small-24B: 81% / 46%; Gemma3-27B: 77% / 42%). In practice, you get near-frontier STEM reasoning without the hardware bill of a flagship model.

Both the 120 B and 20 B checkpoints pass through an extensive safety pipeline: CBRN data filtering, deliberative alignment, and jailbreak testing that matches o4-mini while staying below all “high-risk” thresholds in bio- and cyber-security audits [Research page]. To harden the models further, OpenAI has launched a $500 k Kaggle Red-Teaming Challenge inviting the community to probe gpt-oss-20 B for new vulnerabilities; winning discoveries will feed directly into the public evaluation suite [Kaggle page].

Share

Papers

Chemical Stimulation Sustains Bioluminescence of Living Light Materials [Brachi et al., bioRxiv, July 2025] – IVdB

Why it matters: Haptic sensing remains one of the biggest challenges for robot hands. Looking to nature for inspiration, and subsequently engineering with nature to create engineered living materials capable of performing unique sensing and response functions is an opportunity space to explore further. Being able to detect pressure based on light emitted - for example - whilst also having a material self-heal would be the first wave towards reverse bioaugmentation. Your fingers heal - now just imagine if they luminesced in direct proportion to how strong you interacted with a material!

In this preprint, Brachi et al. set out to explore the use of bioluminescent engineered living materials (ELMs). Part-cell and part-scaffold, ELMs are composites designed to synergistically function, allowing unique sensing, processing, and responding to external environments. What is, and what is not, an ELM? Consider a mussel. A bivalve in the mollusc family, its soft body shelters inside a calcite and protein biomineral composite. Yet whilst attached to its inner shell through adductor muscles, mussels are not ELMs as they are not created by mankind. Furthermore, the shell does not change shape or color upon movement of the mussel; there is no dynamic feedback in the system.

ELMs in contrast do make dynamic feedback loops possible. In this work the concept of an ELM is beautifully presented through a 3D printed dinoflagellate supporting alginate matrix (figure a-b,f). By encapsulating the marine dinoflagellate species Pyrocystis lunula (Pl), Brachi et al. augmented the naturally occurring bioluminescence signal using a second chemical actuator.

Figure 1: a. Diagram of 3D printing workflow and chemical activation of resultant structures. b. From top to bottom, AutoCAD designed structures, printed hydrogels with alginate alone (G), and alginate + cell gels (G+C). c. SEM image of gels. d. Fluorescence image of gels. e. Nonstimulated, and acid and base stimulated PI cells. f. Acid- and base-activated gels. g. Luminescence intensity per condition.

In PI, bioluminescence is mechanically controlled; upon pressure small scintillons - cytoplasmic particles holding luciferase enzyme, luciferin substrate, and luciferin-binding protein - are activated through a signal transduction pathway that lowers the local pH surrounding the scintillon. This proton influx in turn activates the luciferase enzyme, catalysing blue luciferin light emission [1]. Building on the pH responsive change PI exhibits, an acidic (pH 4) and basic (pH 10) condition were evaluated. As shown in figure e, acidic conditions produced high intensity localised flashes within each scintillon. These covered ~4% of each algal body per flash. Basic stimulation in contrast produced localised signals that subsequently diffused out from each scintillon into the wider cytoplasm, covering ~50% of each algal body. Given these different responses, each algal-alginate composite region could be locally controlled by applying either a basic or acidic solution. In effect, patterning or barcoding was possible.

Yet what made this ELM unique was its self-healing nature. Beyond a single barcoding instance, the material was capable of reverting to equilibria. Moreover, by employing an alginate-based matrix, cells were able to proliferate and thrive, expanding colonies over time. In this manner preconditioned acid-based gels could be stimulated mechanically 1x/week over the course of one month, retaining 75% emitted bioluminescence efficiency. For basic-primed gels emission dropped to 0 after 2x of stimulation at 1x/week rate. This was expected given basic conditions ruptured scintillons. Best conditions were thus: acid+force, which emitted 4.46x and 2.12x light as compared to acid or force alone, respectively.

So what level of technology readiness is this at? A: still very much in development (TLR 2-3). And yet, the notion of using ELMs for haptic sensing, self-healing barcodes, or viscoelastically light-emitting responsive packaging is an exciting area to explore further. Curious to see where this technology goes next, and how scintillon stimulation can be made more recurrent and stable over time!

How laypeople evaluate scientific explanations containing jargon [Cruz et al., Nature Human Behavior, June 2025] – PL

Why it matters: Misinformation amplified by social media and answers generated by large language models both have the ability to create incorrect explanations to scientific questions that can satisfy a non-expert. The use of jargon is an important component in how non-experts evaluate answers. The authors aim to understand how jargon impacts the explainability and comprehensibility in answers to scientific questions and ways to correct comprehension by non-experts.

How can we evaluate explanations when doing so requires expertise that we lack? For example, how might a layperson ‘evaluate explanations for the efficacy of COVID-19 mRNA vaccines without substantial knowledge of molecular biology?’ Since this knowledge is increasingly specialized and misinformation more prevalent, the problem has become more pertinent.

To answer this Cruz et al. designed nine experiments with close to 7,000 participants. In these studies, the authors reconcile previously contradictory results: the fact that jargon increases explanatory satisfaction while decreasing comprehensibility.

The authors found that jargon increases explanatory satisfaction for circular explanations, but not for more complete explanations. However, explanations with jargon are judged less comprehensible overall regardless of their completeness. They also propose that jargon inflates perceptions of explanatory satisfaction for poor explanations as when faced with explanatory gaps, people assume that ‘authoritative jargon’ fills those gaps, even when comprehensibility is impaired. Finally, they also found that participants that were asked to generate their own explanations after being presented with circular explanations, acted as a ‘corrective force’.

The Virtual Lab of AI agents designs new SARS-CoV-2 nanbodies [Swanson et al, July 2025] – MB

Why it matters: Agentic AI "co-scientists" for scientific research have generated substantial excitement and skepticism over the past year. Following the release of the AI Virtual Lab paper in Nature this week (original preprint was posted November 2024), we review how the system works and discuss some of the opportunities and challenges with benchmarking AI systems for novel discoveries.

A major question in AI for science is how to best leverage the general-purpose capabilities of large language models (LLMs) for scientific discovery. In "The Virtual Lab of AI agents designs new SARS-CoV-2 nanobodies", researchers at Stanford propose to have an LLM answer user queries by simulating interdisciplinary meetings that may take place in a lab. At its core, the proposed method is simply a prompting strategy: rather than having an LLM respond directly to a user's questions, the LLM is repeatedly prompted to assume specialized roles, such as a principal investigator, scientific critic, immunologist, or computational biologist, and generates a conversation amongst these "agents" (called "meetings") which are synthesized into a response.

The main focus of the paper is an application of the Virtual Lab to design novel antibodies or nanobodies for a recent SARS-COV-2 variant. The authors generated a series of meeting agendas for the Virtual Lab with objectives like project specification, tool implementation, and workflow design . Each stage was scoped fairly narrowly: for example, the project specification prompt is not fully open-ended, but rather instructs the Virtual Lab to decide between designing an antibody or nanobody and whether to generate designs de-novo or through modification (conversations can be viewed on GitHub). Notably, the prompts were engineered specifically for the nanobody design task, and some generated code required additional prompting and intervention to succeed. All together, the Virtual Lab was applied to select 92 nanobody designs modified from existing nanobodies, of which two showed improved binding for the new variant while still maintaining good binding to prior viruses.

The primary strength of this paper is the demonstration of the potential for language models to support a scientifically interesting use case. It may not be immediately obvious that text queries to LLMs can be strung together to solve scientific problems, and one can imagine how agentic methods will become more general and powerful as LLMs improve. Even with imperfect performance from LLMs, scalable tools for scientific reasoning could be highly impactful.

However, the evaluation based on a single proof-of-concept also has significant limitations. Without strong baselines, it is difficult to assess how much the specific proposed solution actually contributes to success. Would a human with similarly specific instructions or an LLM with a different prompting strategy also arrive at similarly successful solutions? While the Virtual Lab paper provides evidence that the multi-role setup leads to responses preferred by users, it is not clear what impact this has on the overall discovery performance or the degree to which the specific design choices are optimal. Given that the result is a one-off success and failures may have not been reported, it is difficult to evaluate how general the proposed system is. These challenges also apply to the evaluation of other AI scientist systems, such as Google's AI co-scientist and Future House's Robin, which highlight key discoveries made by the AI systems as their key results (there has been important critical discussion around the novelty and impact of these discoveries).

Overall, the Virtual Lab highlights the exciting potential of LLMs for augmenting scientific research. As the field moves forward, improved evaluations with strong baselines will be required to develop a clear picture of the real-world utility of current agentic AI co-scientists.

Notable deals

National Science Foundation announces Cornell-led AI Materials Institute

CuspAI targets $100M to unlock AI’s potential in climate materials

In case you missed it

Could AI Slow Science? [Sayash Kapoor and Arvind Narayanan, July 2025]

New NSF Institute for AIxMath [and see Twitter]

What we liked on socials channels

America’s AI Action plan

Terence Tao on the need for AI benchmarking to move from qualitative to quantitative achievements (discussion related to AI IMO results)

Field Trip

Subscribe now

Did we miss anything? Would you like to contribute to Decoding Science by writing a guest post? Drop us a note here or chat with us on Twitter: @pablolubroth @ameekapadia

Welcome to the first issue of Decoding Science! We hope you enjoy our curation of blogs, papers, deals and socials. We’d love to receive your feedback and ideas, especially as we get started. If you have anything to share please email us here.

If you’re interested in covering the Notable Deals section, please message us as well!

Subscribe now

What we read

Blogs

OpenAI and DeepMind achieved gold-medal performance at International Mathematical Olympiad [Decoding Science, July 2025]

We now know there are general-purpose reasoning LLMs that can achieve gold-medal–level performance at the International Mathematical Olympiad (IMO). In 2025, both OpenAI and DeepMind reached this milestone. OpenAI’s experimental reasoning model solved five out of six problems (35/42 points) under exam conditions using only natural language and was independently verified by former IMO medalists. Meanwhile, DeepMind’s Gemini Deep Think became the first model officially certified by the IMO committee to solve five problems with natural-language proofs, leveraging parallel reasoning and reinforcement learning. These achievements mark a turning point, showing that generalist models can now perform at the very frontier of human mathematical ability.

By contrast, in 2024, the top AI performer was DeepMind’s hybrid neuro-symbolic system, combining AlphaProof and AlphaGeometry 2, which scored 28 points, just below the gold threshold (so technically, a silver medal). These were specialized systems that translated IMO problems into formal logic representations and solved them via symbolic search, demonstrating strong performance but requiring significant domain-specific design. This year’s advances represent a shift: from specialized engines to general-purpose LLMs that reason, plan, and solve open-domain problems in natural language. The transition suggests a broader capability trend, one where generalist AI is beginning to rival human experts in deeply structured reasoning domains.

Will it be long before AI achieves a Fields Medal–level breakthrough: proving Fermat’s Last Theorem or solving one of the Millennium Prize Problems, such as the Navier–Stokes equations, as DeepMind hinted earlier this year?

Read more here: DeepMind and OpenAI claim gold in International Mathematical Olympiad [Alex Wilkins, New Scientist, July 2025]

Scientific Publishing: Enough is Enough [Seemay Chou, Astera Institute, June 2025]

The Astera Institute announced that it will no longer fund science that is published through traditional pipelines. Chou motivates this decision by arguing that journals have become “legacy infrastructure,” optimized for prestige and scarcity rather than speed, transparency, or reuse. Moreover, this misalignment shapes what questions scientists ask, how they design experiments, and when they choose to share results. Instead, we need to build new, internet-native, channels for scientific distribution: open-source code and data, make peer feedback a less opaque process, and stop presenting research as a clean narrative.

Such forms of rapid, iterative publishing would let AI agents mine the full arc of a project with negative results, abandoned hypotheses, evolving code notebooks providing richer training data to LLMs that generate hypotheses or design experiments. Transparent science would also make for a more dynamic experience for readers by incentivizing automated tools for scrapping and synthesizing reviews and replications, and employing AI methods to dynamically suggest relevant further reading.

Besides the technical advantages, Chou also highlights the immense financial burden of journal publishing with billions now spent on subscription fees. In contrast, granular, open science could allow funders to redirect resources toward better research infrastructures, FAIR data archives, and prize mechanisms that reward real‑world impact.

This problem is not easily tractable but it raises an interesting possibility – will liberating the scientific record help accelerate discovery?

Share

Papers

The Idea-Execution Gap: Execution Outcomes of LLM-Generated versus Human Research Ideas [Si et al, arXiv, June 2025]

Why it matters: Good benchmarks are crucial for progress in AI: they define our understanding of the current state of the field and guide the objectives that researchers pursue in their work. This paper presents a large-scale study to evaluate the quality of large language model (LLM)-proposed research ideas by actually implementing them and having experts review the quality of the resulting papers. The authors identify an “idea-execution” gap for LLM proposals, which are rated more highly by experts at the proposal stage than after execution. These results emphasize the importance of large sample sizes, strong human baselines, and real-world deployments for making robust claims about the strengths and weaknesses of AI tools for science.

Idea generation is the first step of many "AI scientist" systems. But just how good are AI proposed ideas? A 2024 study by Si et al found that NLP research ideas proposed by large language models (LLM) scored higher on novelty than human proposed ideas as rated by blinded expert reviewers. Last month, Si et al published a follow-up study that investigated whether these higher scores translate into better research projects. They hired 43 expert "execution participants" to implement and write a short paper for randomly assigned ideas (averaging 103 hours per project!) and 58 expert reviewers to score the generated papers. All participants were blinded as to whether each idea was proposed by a human or an LLM. Interestingly, the authors found that scores for AI generated ideas drop substantially from the proposal to execution stage, which they refer to as the "idea-execution gap" for LLM generated ideas. By contrast, human proposed ideas were scored similarly before and after execution. The authors attribute the difference in scores to a greater emphasis on feasibility and empirical performance in reviews after execution.

Limitations: The idea generation process in this study differs from standard research: ideas are fixed at the beginning of each project, and problems are constrained to innovations in LLM prompting techniques, limiting opportunities for creativity. Additionally, the LLM proposed ideas were generated almost a year ago, so LLM improvement may impact the results if rerun today. Finally, I suspect that the peer-review score evaluation metric may carry a risk of a "publication-utility gap" even after project execution (publication criteria are unfortunately not always aligned with real-world utility), and it may be valuable for future work to explore idea generation metrics that are grounded in the intended impact of the research.

3D Atomic-Scale Metrology of Strain Relaxation and Roughness in Gate-All-Around (GAA) Transistors via Electron Ptychography [Karapetyan et al., arXiv, July 2025]

Why it matters: Reconstructing Gate-All-Around (GAA) devices in a simulated model with atomic precision can surface defects and variations that dictate carrier mobility and leakage at sub-3 nm scales. In creating these single-atom precise models GAA architectures can be developed with faster iteration speed and minimal fabrication trial-and-error, thereby accelerating the transition to ultra-dense energy-efficient semiconductor devices.

Field-effect transistors, particularly metal-oxide semiconductor FETs (MOSFETs) are the fundamental layer enabling virtually all modern digital interactions. Improvements in FETs have made everything from pocket-sized computers (smartphones) to AI-enabled cloud infrastructure possible. Shifting from 2D FETs and FinFETs into 3D Gate-All-Around (GAA) devices has made shorter gate lengths (2-3 nm) possible without increased leakage, effectively continuing to push Moore’s scaling law further. Vertical fabrication of nanosheets and nanowires presents new opportunities to tune 3D architectures and increase logic density, achieved through higher transistor count within the same area.

However, fabrication of nanoscale features in 3D remains an intricate and complex task requiring angstrom-scale precision. Moreover, if new device architectures are to be designed to maximize carrier mobility (the speed at which electrons, or holes*, move through semiconductor devices when an electric field is applied), understanding how to optimize fabrication is essential.

To this end, Karapetyan et al. give a new 3D perspective into semiconductor architectures. Resolving features at 0.49 A width, multislice electron ptychography (MEP) surpasses previous techniques at 0.66 A (tf-iDPC) and 0.83 A (tf-ADF). Figure a-d presents how MEP operates, and what data can be produced. In achieving this greater resolution, it enables uncovering stacking faults (white lines in g), step edges (green arrows in g), interface roughness, pinholes, and strain-relaxations each critical to defining carrier mobility. Beyond imaging, MEP is an interesting technique due to its capacity to reliably reconstruct atomic potentials. Effectively, this means the technique can both image and provide information about the electronic field present. Furthermore, MEP can i) image and reconstruct an atom-scale boundary through a single scan, and ii) resolve features at 40 nm depth (this paper showed 38 nm). Lack of electron scattering further enhances quality of reconstructed models, and predictability of simulated structures.

So what are the next steps? In defining strain-relaxation as a key metric for interface quality, Karapetyan et al. could start to accelerate the speed at which new GAA devices are designed. At present, it still takes 3-4 months, and > 1000 steps!, to form the critical GAA structure that enables superior device performance. Building an atomic-scale model from 3D reconstructed characterization results could improve processing conditions and give new insights into fabrication techniques. And whilst the 2-3 nm limit remains a physical constraint defined by electron tunneling, a shift from 2D to 3D opens up the possibility to shift the semiconductor industry into a new era of vertically integrated, energy efficient architectures with unprecedented density and design flexibility.

*holes = empty states in the valence band that appear when an electron hops into the conduction band. They effectively describe the space that is left behind, and have a charge equal and opposite (+e) to that of an electron (-e).

Model-based reinforcement learning for ultrasound-driven autonomous microrobots [Medany et al., Nature Machine Intelligence, June 2025]

Why it matters: The study is a proof‑of‑concept that model‑based RL can solve a millisecond‑scale control problem. By fusing Dreamer v3 model‑based reinforcement learning with fast image tracking, the team shows that its policy can transfer from simulation and adapt in minutes, solving a millisecond‑scale steering problem. This cuts experimentation time by orders of magnitude and opens the door to microrobots that can autonomously navigate blood vessel-sized channels.

This group from ETH and IBM investigated the control of ultrasound-driven microrobots through model-based reinforcement learning (MBRL). MBRL is a group of RL methods that learn a model of the environment and then use that model to refine a policy, instead of relying purely on trial and error. These microrobots have emerged as a non-invasive alternative, capable of generating tunable propulsive forces, enabling deep navigation into tissues. However, achieving precise control of these microrobots is challenging as several transducers need to be controlled with millisecond resolution to effectively steer; which is too complex for human operators.

To allow autonomous control of ultrasound-driven microrobots, the authors employed Dreamer v.3 MBRL. MBRL is an RL strategy that learns a world model of the microrobot's environment and imagines the future attempts inside the model to plan actions, instead of only relying on trial and error. Image-processing techniques were used to track and detect swarms in real time, framing the control over the microrobots as an RL task. Dreamer v.3 was trained within an imagined model to simulate and imagine future states in the environment. A simulated game environment was designed to pretrain the microrobots to reduce the convergence time during experimental training.

The experimental set-up included an artificial vascular channel encircled by eight piezo-transducers (PZTs). This was all mounted on an inverted microscope to capture the results. The microrobots are produced by the self-organization of microbubbles in an ultrasound field. The PZTs act as the steering mechanism, which generates a pressure gradient for the microrobots to move away from the activated transducer in a perpendicular direction from the PZT. Each action includes the choice of one PZT and its frequency and amplitude. The large continuous action space is particularly favoured by MBRL. Against Proximal Policy Optimization (PPO), Dreamer v.3 converged 50x faster across all virtual channel geometries. A policy trained in the simulation hit 50% target-reach success and climbed to 90% within 30 mins of on-line fine-tuning in new channels.

Notable deals

Diode Computers raises $11.4M led by a16z. Diode is embedding AI into the design process of circuit boards to generate boards that are functional and manufacturable at scale.

Israeli-based quantum computing startup Qedma raises $26M from IBM & others for its noise-resilient software.

Denmark’s Export and Investment Fund, EIFO, invests $5M in defence technology fund D3, strengthening the EU <> Ukraine connection and access to cutting edge defence tech. They are the first state-backed fund to invest in D3, with previous D3 backers including Eric Schmidt, among others.

EFFECT Photonics raises an additional $24M as part of its series D round, to continue building photonic chips computing with light (rather than electricity).

Robinhood CEO AI Math Start-Up Valued at Nearly $900M. Harmonic AI’s founder, Vlad Tenev, has raised $100M in funding to tackle a problem that has sometimes confounded AI models: math.

In case you missed it

Aeneas transforms how historians connect the past [Google DeepMind, July 2025]

What we liked on socials channels

Field Trip

Subscribe now

Did we miss anything? Would you like to contribute to Decoding Science by writing a guest post? Drop us a note here or chat with us on Twitter: @pablolubroth @ameekapadia

Artificial intelligence and automation is transforming every scientific discipline: from physics and chemistry to materials science, earth systems, and space exploration.

AI is not only accelerating breakthroughs; it’s reshaping the scientific process through automated hypothesis generation, experiment design optimization, streamlining literature synthesis, and next-generation lab automation.

We’ve started Decoding Science to track and explain the most important papers, tools, deals, funding models, and other developments in how AI is contributing to advances in everything from quantum physics, chemical synthesis, electronic engineering, self-driven labs and material discovery to science itself.

Our biweekly newsletter curates and explains the most exciting recent papers, tools, platforms, and news at the AI in Science interface.

Subscribe now

Our incredible team of authors that will cover this diverse set of topics include:

• Hiya Jain: Neuroscience and History @ Columbia University. Emergent Grants Fellow. Covering metascience and the history of science.

• Inga Van den Bossche: DPhil in Engineering/Biomaterials @ Oxford, EVP of Research at Nucleate. Covering novel material science.

• Michael Bereket: AI for Science, CS PhD candidate @ Stanford. Covering the role of AI in the scientific process.

If you’re passionate about the future of science and believe clear storytelling can shape its trajectory, join us.