Search-R1

Trains LLMs with RL (PPO/GRPO) to interleave step-by-step reasoning with live search-engine calls — multi-turn retrieval inside the rollout, retrieved-token masking for stable optimization, and up to ~26% (Qwen2.5-7B) gains over RAG baselines across seven QA datasets. From Bowen Jin in Jiawei Han's group (with UMass and Google Cloud AI), it became the canonical baseline of the 2025 search-RL / deep-research-agent wave: DeepResearcher, VRAG-RL, and most successors compare against it.