Xiangqing Shen

FARCA transforms factual supervision into precise, reliability-weighted training signals, significantly boosting model factuality without sacrificing reasoning performance.

Forget static relevance labels – RRPO uses LLM feedback to train RAG rerankers, boosting generation quality without expensive human annotations.

Ditch the costly process supervision: EVOM trains LLMs to generate optimization code by treating solvers as deterministic reward functions, enabling zero-shot solver transfer and efficient adaptation.