Paul Pu Liang

We present MimeQA, a question-answering dataset on mime videos, to evaluate video LLMs' nonverbal social reasoning capabilities, and found that models perform below human performance.

The paper proves that a two-layer, single-head transformer can reliably perform in-context learning on any-order Markov chains.

We built a generalist clinical foundation model across both time series and vision modalities with a novel RL training algorithm named DRPO.

We argue that simulating society with generative agents requires moving beyond behavioral mimicry toward cognitively grounded reasoning that is structured, revisable, and traceable.