Haoyi Qiu

I am a second-year PhD candidate in Computer Science at UCLA advised by Prof. Nanyun (Violet) Peng. Before, I graduated from the University of Michigan, with a B.S. in Computer Science and Math, advised by Prof. Joyce Y. Chai.

My research addresses fundamental questions in model evaluation, safety alignment, multimodal understanding, and multimodal reasoning:

• How can we rigorously evaluate AI systems (LLMs, VLMs, agents, diffusion models) beyond surface-level metrics?
• How do we align foundation models with diverse cultural norms and ensure geo-diverse safety through post-training?
• How do we ensure foundation models understand and respect pluralism?
• How can we build models that reason effectively across modalities while remaining trustworthy and aligned with human values?

I have been recognized with the Outstanding Graduate MS Student Award from UCLA CS Department (2025) and the Outstanding Reviewer Award at EMNLP 2023 (top 0.74%).

If you're interested in my research, potential collaborations, or simply want to catch up, feel free to drop me an email.

Email: [email protected] 💌 /  Google Scholar 👩🏻‍🎓  /  Twitter 📝  /  Life 🌷  / 

* denotes Equal Contributions and Project Lead

We introduce VideoBiasEval, a diagnostic framework that traces how social biases evolve through the alignment pipeline of video diffusion models. By training our own reward models and alignment-tuned diffusion models, we provide a full training recipe and reveal that alignment tuning not only strengthens representational biases but also makes them temporally stable across generated videos.

We propose MMGR, a principled evaluation framework for generative reasoning across abstract reasoning, embodied navigation, and physical commonsense, revealing that leading video and image models struggle with causal correctness and long-horizon spatial planning.

We introduce CROSS, a benchmark of 1,284 multilingual visually grounded queries from 16 countries to assess cultural safety in LVLMs. We develop supervised fine-tuning with culturally grounded data and preference tuning with contrastive response pairs, substantially improving GPT-4o's cultural awareness (+60%) and compliance (+55%) while preserving general capabilities.

We introduce MMPersuade, a unified framework for studying multimodal persuasion in LVLMs, showing that multimodal inputs substantially increase model susceptibility to persuasion—especially in misinformation scenarios—and that different persuasion strategies vary in effectiveness across contexts.

We propose GUI-KV, a plug-and-play KV cache compression method for GUI agents that exploits spatial saliency and temporal redundancy, reducing decoding FLOPs by 38.9% while increasing step accuracy by 4.1% over full-cache baselines without retraining.

We propose CogAlign, a post-training strategy inspired by Piaget's cognitive development theory that trains VLMs to recognize invariant properties under visual transformations. CogAlign outperforms supervised fine-tuning while requiring 60% less training data, significantly improving visual arithmetic and downstream chart and geometry understanding.

We introduce CASA, a benchmark for evaluating LLM web agents' cultural and social awareness across shopping and forum tasks, finding that agents achieve less than 10% awareness coverage with over 40% norm violation rates. We show that fine-tuning on culture-specific datasets enhances cross-region generalization, while prompting boosts complex task navigation.

We provide a comprehensive survey of automatic chart understanding in the era of large foundation models, covering tasks, evaluation metrics, modeling strategies, and future directions including domain-specific charts and agent-oriented settings.

We introduce SafeWorld, a benchmark of 2,342 queries grounded in human-verified cultural norms and legal policies from 50 countries, along with a multi-dimensional safety evaluation framework. Using synthesized preference pairs for DPO, our trained SafeWorldLM outperforms GPT-4o across all evaluation dimensions by a large margin.

We propose BiasPainter, an evaluation framework that automatically triggers and detects social biases in image generation models across 62 professions, 39 activities, 57 objects, and 70 personality traits, achieving 90.8% accuracy on automatic bias detection.

We introduce a multi-dimensional benchmark and LLM-based evaluation framework for LVLM hallucination that covers objects, attributes, and relations, generalizing the CHAIR metric to jointly assess both faithfulness and coverage of model outputs.

We propose AMRFact, a framework that leverages Abstract Meaning Representations to generate coherent factually inconsistent summaries with high error-type coverage for training factual consistency evaluators, significantly outperforming previous systems on the AggreFact-SOTA benchmark.

We conduct a systematic study of gender biases in model-based evaluation metrics (e.g., CLIPScore) for image captioning, revealing that using biased metrics as rewards in RL training propagates and amplifies stereotypes in generation models. We propose an effective debiasing method without hurting correlation with human judgments.

Outstanding Graduate MS Student (1 per department) - UCLA CS Department, 05/2025

Outstanding Reviewer Award (0.74%) - EMNLP 2023

James B. Angell Scholar, University Honors, EECS Scholar, The University of Michigan

Area Chair: ACL Rolling Review 2025

Reviewer: ICLR 2025, ACL Rolling Review 2024, EMNLP 2023

This website is built using the source code from Jon Barron's public academic website.