Yasheng Wang

We propose a new distillation approach that removes the input question for adaptive and efficient reasoning.

Our paper introduces WebPuzzle, a novel dataset boosting LLMs' real-world info-seeking capability, and DeepDiver, an RL-based framework enabling dynamic Search Intensity Scaling for iterative evidence gathering.

MMDocRAG, a comprehensive multimodal DocRAG benchmark

RidgeLoRA adopts a novel series connection architecture, and enables different parameter initialization approaches for better performances.