{"paper":{"title":"DeepResearcher: Scaling Deep Research via Reinforcement Learning in Real-world Environments","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"End-to-end RL training on the open web lets LLM agents outperform prompt and RAG baselines by up to 28.9 points while developing planning and self-reflection.","cross_cats":["cs.CL","cs.LG"],"primary_cat":"cs.AI","authors_text":"Dayuan Fu, Lyumanshan Ye, Pengfei Liu, Pengrui Lu, Xiangkun Hu, Xiaojie Cai, Yuxiang Zheng","submitted_at":"2025-04-04T04:41:28Z","abstract_excerpt":"Large Language Models (LLMs) equipped with web search capabilities have demonstrated impressive potential for deep research tasks. However, current approaches predominantly rely on either manually engineered prompts (prompt engineering-based) with brittle performance or reinforcement learning within controlled Retrieval-Augmented Generation (RAG) environments (RAG-based) that fail to capture the complexities of real-world interaction. In this paper, we introduce DeepResearcher, the first comprehensive framework for end-to-end training of LLM-based deep research agents through scaling reinforce"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"DeepResearcher achieves substantial improvements of up to 28.9 points over prompt engineering-based baselines and up to 7.2 points over RAG-based RL agents, with emergent cognitive behaviors including planning, cross-validation, self-reflection, and honesty.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"That the multi-agent browsing architecture can reliably extract information from arbitrary real-world webpage structures at scale without introducing systematic biases or instability that would undermine the reported performance gains.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"End-to-end RL in authentic web environments produces LLM research agents that outperform prompt-engineering and RAG-based baselines by up to 28.9 and 7.2 points respectively while exhibiting emergent planning, cross-validation, and self-reflection.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"End-to-end RL training on the open web lets LLM agents outperform prompt and RAG baselines by up to 28.9 points while developing planning and self-reflection.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"eea7bbcef4bd1c25a2bd03aed38de86b097ddc4e7746d2e24b0d8f87cf649252"},"source":{"id":"2504.03160","kind":"arxiv","version":4},"verdict":{"id":"e8c3be02-a679-4ccc-b3f5-bf3cc6a75514","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-16T19:55:39.109853Z","strongest_claim":"DeepResearcher achieves substantial improvements of up to 28.9 points over prompt engineering-based baselines and up to 7.2 points over RAG-based RL agents, with emergent cognitive behaviors including planning, cross-validation, self-reflection, and honesty.","one_line_summary":"End-to-end RL in authentic web environments produces LLM research agents that outperform prompt-engineering and RAG-based baselines by up to 28.9 and 7.2 points respectively while exhibiting emergent planning, cross-validation, and self-reflection.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"That the multi-agent browsing architecture can reliably extract information from arbitrary real-world webpage structures at scale without introducing systematic biases or instability that would undermine the reported performance gains.","pith_extraction_headline":"End-to-end RL training on the open web lets LLM agents outperform prompt and RAG baselines by up to 28.9 points while developing planning and self-reflection."},"references":{"count":18,"sample":[{"doi":"","year":2025,"title":"Qwen, :, An Yang, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chengyuan Li, Dayiheng Liu, Fei Huang, Haoran Wei, Huan Lin, Jian Yang, Jianhong Tu, Jianwei Zhang, Jianxin Yang, Jiaxi ","work_id":"e84e2134-b769-4027-8225-600d73ff84a9","ref_index":1,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2023,"title":"Timo Schick, Jane Dwivedi-Yu, Roberto Dess`ı, Roberta Raileanu, Maria Lomeli, Eric Hambro, Luke Zettle- moyer, Nicola Cancedda, and Thomas Scialom. 2023. Toolformer: Language models can teach themselv","work_id":"6fd18ea3-e439-4393-b21e-7dbddf8d3dd4","ref_index":2,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2025,"title":"R1-Searcher: Incentivizing the Search Capability in LLMs via Reinforcement Learning","work_id":"f5ed73d2-f2ff-4cf6-853d-3586333e44ef","ref_index":3,"cited_arxiv_id":"2503.05592","is_internal_anchor":true},{"doi":"","year":2025,"title":"Kimi k1.5: Scaling Reinforcement Learning with LLMs","work_id":"bff96ab1-bd6a-4585-be23-74fdb51969c7","ref_index":4,"cited_arxiv_id":"2501.12599","is_internal_anchor":true},{"doi":"","year":2022,"title":"Harsh Trivedi, Niranjan Balasubramanian, Tushar Khot, and Ashish Sabharwal. 2022. MuSiQue: Multihop questions via single-hop question composition. Transactions of the Association for Computational Lin","work_id":"9a1f5b9e-1481-4ef8-9826-dda294612579","ref_index":5,"cited_arxiv_id":"","is_internal_anchor":false}],"resolved_work":18,"snapshot_sha256":"7eb220dfde26546444931055c723153fc67b25c640face9c0c5a4787d50e602a","internal_anchors":2},"formal_canon":{"evidence_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"author_claims":{"count":0,"strong_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"builder_version":"pith-number-builder-2026-05-17-v1"}