Yin Li

Streaming VLMs can now internalize long-term history as evolving LoRA weights, bypassing context window limitations and achieving superior performance on video understanding tasks.

Jointly optimizing visual token selection and LLM computation can drastically reduce inference costs while enhancing performance in multimodal tasks.

TOFFEE can synthesize high-quality data agent trajectories that significantly improve LLM performance in unfamiliar analytical workflows.

Current vision-language models falter in streaming interaction understanding, with alarming mis-calibration leading to confidently incorrect predictions.

Unlocking video generation in any temporal order, UniTemp enables innovative workflows like bidirectional extensions and visual story generation without sacrificing performance.

DRIFT transforms the generative modeling landscape by simplifying the optimization of continuous outputs, leading to superior performance in both perception and robotic control tasks.