Cog Lunch: Wei-Chen (Eric) Wang (Seethapathi Lab) & Jeonghwan (Jay) Lee (Seethapathi Lab)

May 12, 2026
12pm
Location: 46-3189

Zoom: https://mit.zoom.us/j/94900932731

Speaker: Jeonghwan (Jay) Lee

Affiliation: Seethapathi Lab
Title: How Human-Like is Generative AI for Human Movement?

Abstract: Motor behavior offers a uniquely demanding test for generative AI: it is high-dimensional, temporally extended, and requires both feedforward planning and reactive feedback control. We evaluate three state-of-the-art generative AI models of human movement against quantitative signatures of feedforward and feedback human motor control. We find that neither the biological fidelity of the body model nor the scale and diversity of training data guarantee human-like behavior. A simpler model trained to match the statistics of real human state transitions, rather than specific target poses or a latent motion distribution, outperforms both, suggesting that how the learning signal is grounded in movement data matters more than body complexity or data breadth. These findings have direct implications for building generative models of movement that are faithful enough to serve as predictive simulations, whether for understanding motor control or designing technologies that interact with the human body.

Speaker: Wei-Chen (Eric) Wang

Affiliation: Seethapathi Lab
Title: Aligning Generative AI with Fewer Preference Queries

Abstract: Generative AI can now produce a vast repertoire of human movement. Aligning that repertoire to an individual, thereby modeling how they choose to move and why, would enable personalized movement prediction with applications to neuromotor rehabilitation and wearable robotics. However, this alignment relies on preference feedback, which is severely limited in practice: evaluating movement is cognitively demanding, and reliable judgments are exhausted after only a few dozen comparisons. The "manifold hypothesis" offers a principled escape: meaningful behavioral variation around any given movement is locally low-dimensional. We show that our algorithm, COMPASS, significantly outperforms standard preference learning algorithms, and succeeds on tasks where all baselines fail to improve over random search. Restricting the search to this neighborhood transforms an intractable global problem into a tractable local one.