Sheng Chen

Standard accuracy metrics can fool you: LLMs often appear strong at SAT solving, but a paired-formula test reveals they're mostly guessing "satisfiable" and failing to actually reason.

Recommender systems can move beyond passive item lists: RecPilot's multi-agent framework autonomously explores item spaces and generates user-centric reports, significantly reducing user effort in item evaluation.

Finally, a single model handles multi-modal video generation, inpainting, and editing at cinematic resolutions with synchronized audio, all while accepting diverse inputs like text, images, video clips, and audio references.

LLMs learn to recommend better by looking inside themselves, using intermediate layer activations to generate harder negatives on the fly.