Shengyao Zhuang

\textsc{Sieve} achieves higher accuracy with significantly fewer tokens by intelligently filtering webpage content based on structure, revolutionizing how deep-research agents retrieve information.

Achieving high-quality reranking with a 30B MoE model is now feasible on an academic budget, outperforming traditional dense models in efficiency.

MPE achieves superior long-context retrieval by efficiently encoding document chunks while maintaining critical contextual relationships, outperforming traditional methods.

Forget scaling and RLHF: carefully selecting internal attention signals from the right layers lets a zero-shot 8B model match a 14B reinforcement-learned re-ranker in complex reasoning tasks.