Kai Ye

Lucida redefines scene modeling by achieving 69% better accuracy than existing methods while simplifying the data requirements at each processing step.

Test-time scaling can significantly enhance LLM reasoning capabilities, but without clear protocols, results are often incomparable and misleading.

Token-level detection reveals the specific contributions of LLMs in collaborative writing, outperforming traditional document-level methods.

Single-rollout RL can rival multi-rollout performance for LLM reasoning, thanks to a new batchwise advantage estimation technique that dramatically improves value function accuracy.

Reasoning beats scale: a 1.5B parameter model, READER, outperforms models 100-1000x larger in detecting AI-generated text by explicitly generating a rationale for its decision.