Keyu He (何柯宇) [Kuh-yoo Huh]

Current evaluations for LLMs in multi-agent social interaction settings, where agents must negotiate and collaborate, lag behind other domains with objective ground-truth (e.g. math, logic), instead relying on human ratings or LLM-as-judge scoring, which are subjective, costly, and hard to verify at scale. To address this evaluation gap, we introduce Social Gym, an environment of 21 multi-agent social games where agent performance is verifiable and objective (e.g., Werewolves, Resistance, Spyfall, etc.) organized into five categories along the information-and-communication axes, with a Elo tournament that produces a cross-game leaderboard. Motivated by benchmarking experiments showing limitations of current LLM agents, we propose SPaRTan (Self-Play and Reflect-Transfer), a training-free self-improvement loop where a model plays a game, reflects on its trajectories to produce a transferable playbook, and applies that playbook in subsequent games. Our experiments on GPT-5-mini show that agents armed with a SPaRTan playbook gain on the structurally weaker side of asymmetric hidden-role games, and that the same playbook transfers across games and into weaker student models; the effect is capacity-dependent, however: an open-weights replication with Qwen3-32B is largely null except on Prisoner’s Dilemma. Together, Social Gym and SPaRTan offer a reproducible, verifiable foundation for measuring and improving LLM social reasoning without weight updates.

Understanding how developers interact with AI coding assistants requires more than chat logs or git histories in isolation; it requires reconstructing the full context: which prompt led to which edit, what the developer tried and discarded, and how their strategy evolved over time. We present RECAP (Replay and Examine Captured AI Programming), an open-source platform that (1) passively records AI chat sessions and fine-grained code edits inside VS Code without disrupting the developer’s workflow, (2) merges them into a unified timeline for interactive session replay, and (3) provides extensible analysis modules for behavioral classification and AI reliance measurement. Deployed in a university software engineering course, RECAP captured 2,034 prompts and 8,239 code edits from 41 students across a multi-week project. We demonstrate how the platform’s linked data and replay capabilities enable analyses of developer-AI interaction patterns that no single data source could support.

When people query Vision-Language Models (VLMs) but cannot see the accompanying visual context (e.g. for blind and low-vision users), augmenting VLM predictions with natural language explanations can signal which model predictions are reliable. However, prior work has found that explanations can easily convince users that inaccurate VLM predictions are correct. To remedy undesirable overreliance on VLM predictions, we propose evaluating two complementary qualities of VLM-generated explanations via two quality scoring functions. We propose Visual Fidelity, which captures how faithful an explanation is to the visual context, and Contrastiveness, which captures how well the explanation identifies visual details that distinguish the model’s prediction from plausible alternatives. On the A-OKVQA, VizWiz, and MMMU-Pro tasks, these quality scoring functions are better calibrated with model correctness than existing explanation qualities. We conduct a user study in which participants have to decide whether a VLM prediction is accurate without viewing its visual context. We observe that showing our quality scores alongside VLM explanations improves participants’ accuracy at predicting VLM correctness by 11.1%, including a 15.4% reduction in the rate of falsely believing incorrect predictions. These findings highlight the utility of explanation quality scores in fostering appropriate reliance on VLM predictions.

Language models today are widely used in education, yet their ability to tailor responses for learners with varied informational needs and knowledge backgrounds remains under-explored. To this end, we introduce ELI-WHY, a benchmark of 13.4K "Why" questions to assess the pedagogical capabilities of LLMs. We then conduct two extensive human studies to assess the utility of LLM-generated explanatory answers (explanations) on our benchmark, tailored to three distinct educational grades: elementary, high-school, and graduate school. In our first study, human raters assume the role of an "educator" to assess model explanations’ fit to different educational grades. We find that GPT-4-generated explanations match their intended educational background only 50% of the time, compared to 79% for human-curated explanations. In our second study, human raters assume the role of a learner to assess if an explanation fits their own informational needs. Results show that users deemed GPT-4-generated explanations relatively 20% less suited to their informational needs, particularly for advanced learners. Additionally, automated evaluation metrics reveal that GPT-4 explanations for different informational needs remain indistinguishable in their grade-level, limiting their pedagogical effectiveness. These findings suggest that LLMs’ ability to follow inference-time instructions alone is insufficient for producing high-utility explanations tailored to users’ informational needs.

In open-ended generative tasks like narrative writing or dialogue, large language models often exhibit cultural biases, showing limited knowledge and generating templated outputs for less prevalent cultures. Recent works show that these biases may stem from uneven cultural representation in pretraining corpora. This work investigates how pretraining leads to biased culture-conditioned generations by analyzing how models associate entities with cultures based on pretraining data patterns. We propose the MEMOed framework (MEMOrization from pretraining document) to determine whether a generation for a culture arises from memorization. Using MEMOed on culture-conditioned generations about food and clothing for 110 cultures, we find that high-frequency cultures in pretraining data yield more generations with memorized symbols, while some low-frequency cultures produce none. Additionally, the model favors generating entities with extraordinarily high frequency regardless of the conditioned culture, reflecting biases toward frequent pretraining terms irrespective of relevance. We hope that the MEMOed framework and our insights will inspire more works on attributing model performance on pretraining data.