Feng Chen

EviSelect achieves a 3.9x speedup in long video understanding by dynamically selecting relevant frames based on the MLLM's internal attention evidence, cutting visual token selection by half.

Evolving agents can achieve up to a 19.37-point improvement in task performance by effectively leveraging prior experience in financial workflows.

Role-Grounded Rubric Construction reveals that professional deliverables are essential for accurately evaluating specialized financial AI agents, outperforming traditional prompt-based methods.

Despite advances in LLMs, they fail to effectively integrate user preferences over time, with accuracy rates stagnating around 39% even in ideal conditions.

STACE reveals that autonomous stress-testing with LLM agents can dramatically enhance the evaluation of concept erasure, outperforming traditional methods in robustness and adaptability.

Teaching emotional support chatbots specific, executable skills, rather than relying on end-to-end training, leads to more interpretable, controllable, and ultimately more helpful conversations.