{"paper":{"title":"MMSearch-R1: Incentivizing LMMs to Search","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"Reinforcement learning lets multimodal models search the internet only when needed","cross_cats":["cs.CL"],"primary_cat":"cs.CV","authors_text":"Bo Li, Bo You, Jinming Wu, Wei Li, Yiding Liu, Zejun Ma, Zihao Deng, Ziwei Liu","submitted_at":"2025-06-25T17:59:42Z","abstract_excerpt":"Robust deployment of large multimodal models (LMMs) in real-world scenarios requires access to external knowledge sources, given the complexity and dynamic nature of real-world information. Existing approaches such as retrieval-augmented generation (RAG) and prompt engineered search agents rely on rigid pipelines, often leading to inefficient or excessive search behaviors. We present MMSearch-R1, the first end-to-end reinforcement learning framework that enables LMMs to perform on-demand, multi-turn search in real-world Internet environments. Our framework integrates both image and text search"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"We present MMSearch-R1, the first end-to-end reinforcement learning framework that enables LMMs to perform on-demand, multi-turn search in real-world Internet environments... our model not only outperforms RAG-based baselines of the same model size, but also matches the performance of a larger RAG-based model while reducing search calls by over 30%.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"The outcome-based reward combined with a search penalty, together with the curated search-balanced dataset, is sufficient to produce efficient on-demand search behavior that generalizes beyond the training distribution.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"MMSearch-R1 uses reinforcement learning to train multimodal models for on-demand multi-turn internet search with image and text tools, outperforming same-size RAG baselines and matching larger ones while cutting search calls by over 30%.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"Reinforcement learning lets multimodal models search the internet only when needed","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"3447c2a81aad5ee45d9eec2cce236f98202eda62d8441b689c80c1450340ee79"},"source":{"id":"2506.20670","kind":"arxiv","version":1},"verdict":{"id":"18f9ea2e-53a7-4c65-865b-759ce7bedc8c","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-16T15:21:56.253302Z","strongest_claim":"We present MMSearch-R1, the first end-to-end reinforcement learning framework that enables LMMs to perform on-demand, multi-turn search in real-world Internet environments... our model not only outperforms RAG-based baselines of the same model size, but also matches the performance of a larger RAG-based model while reducing search calls by over 30%.","one_line_summary":"MMSearch-R1 uses reinforcement learning to train multimodal models for on-demand multi-turn internet search with image and text tools, outperforming same-size RAG baselines and matching larger ones while cutting search calls by over 30%.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"The outcome-based reward combined with a search penalty, together with the curated search-balanced dataset, is sufficient to produce efficient on-demand search behavior that generalizes beyond the training distribution.","pith_extraction_headline":"Reinforcement learning lets multimodal models search the internet only when needed"},"references":{"count":94,"sample":[{"doi":"","year":2025,"title":"Open deep search: Democratizing search with open-source reasoning agents","work_id":"467e5b5b-a169-4539-8ddc-bd34e07b0e3a","ref_index":1,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2024,"title":"Claude 3.5 Sonnet","work_id":"45d58c07-b237-4bb8-ac8d-8c68f4784507","ref_index":2,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2023,"title":"Self-rag: Learn- ing to retrieve, generate, and critique through self-reflection","work_id":"cd9884d5-53dc-4ba6-98d6-a38b02cf052e","ref_index":3,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2024,"title":"Mint-1t: Scaling open-source multimodal data by 10x: A multimodal dataset with one trillion tokens","work_id":"9350a215-8bdd-4a04-a217-295413298f83","ref_index":4,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2025,"title":"Qwen2.5-VL Technical Report","work_id":"69dffacb-bfe8-442d-be86-48624c60426f","ref_index":5,"cited_arxiv_id":"2502.13923","is_internal_anchor":true}],"resolved_work":94,"snapshot_sha256":"2a80b22b2e8323d5b48d357cc8e67cc82540c731f2a5ec48516d7e6cffe9fca1","internal_anchors":26},"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"}