A Bayesian cognitive model of epistemic language understanding, which translates natural language into an epistemic language-of-thought (ELoT), then evaluates these translations against inferences about an agent's mental states produced by inverse planning with a Bayesian theory of mind (BToM).

This dataset and model were largely authored by Lance Ying and Tan Zhi-Xuan.

See the paper for more details:

Lance Ying*, Tan Zhi-Xuan*, Lionel Wong, Vikash Mansinghka, and Joshua B. Tenenbaum. 2025. “Understanding Epistemic Language with a Language-augmented Bayesian Theory-of-Mind,” in Transactions of the Association for Computational Linguistics (TACL).

Our LaBToM model comprises two interlinked modules:

• The first module (top), an epistemic language of thought (ELoT), models our capacity to compositionally represent the world (including the contents of others' minds), and how we flexibly translate such thoughts from natural language.

• The second module (bottom), a Bayesian theory-of-mind (BToM), captures our intuitive inferences about others' minds via Bayesian inference over a generative model of how agents update their beliefs and act towards their goals.

Epistemic language understanding can thus be modeled by mapping language into ELoT formulas, which we evaluate against the inferences produced by rational mentalizing.

To set up the environment for this project, make sure the epistemic_language directory is set as the active environment. In VSCode with the Julia for VSCode extension, you can do this by right-clicking the directory in the file explorer, and selecting Julia: Activate This Environment. Alternatively, you can use the activate command in the Julia REPL's Pkg mode.

Then run the following commands in via Pkg mode in the Julia REPL:

add https://github.com/probcomp/InversePlanning.jl.git#release-0.0.2
instantiate

Several scripts in this repository (e.g. threshold fitting, analysis, and plotting code) require access to human-provided data. This data can be downloaded from the files tab of https://osf.io/xq9c4/, and placed in results/humans directory.

• The dataset directory contains all plans, problems, and stimuli
• The src directory contains non-top-level source files
• The scripts directory contains the main experiment scripts
• The translation directory contains ELoT translation code (see translation/README.md)
• The llm_baselines directory contains LLM baselines (see llm_baselines/README.md)
• The assets directory contains image assets used in this README

In the scripts directory, you will find the following scripts:

• run_btom_inference.jl generates and saves BToM inference results via Bayesian inverse planning
• run_elot_evaluation.jl evaluates ELoT translations against saved BToM inference results
• run_elot_evaluation_full.jl evaluates ELoT translations for the full dataset of human-written statements
• run_inference_and_evaluation.jl runs the BToM and ELoT modules consecutively, without intermediary files
• run_threshold_fitting.jl fits threshold parameters for the ELoT module against human data
• testbed.jl is for experimenting with modeling and inference parameters
• stimuli.jl generates stimuli animations and metadata

Scripts for analysis and plotting can also be found in the scripts directory:

• analysis.jl generates analysis results
• plot_statistics.jl generates statistical plots from the analysis results
• plot_storyboards.jl generates storyboards with results for each scenario/stimulus