{"paper":{"title":"LMM-R1: Empowering 3B LMMs with Strong Reasoning Abilities Through Two-Stage Rule-Based RL","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"Strengthening reasoning first on text data then transferring to images improves 3B multimodal models without extra multimodal data.","cross_cats":["cs.AI"],"primary_cat":"cs.CL","authors_text":"Gongrui Zhang, Jie Liu, Kai Yang, Miaosen Zhang, Qipeng Zhu, Xin Geng, Xingzhong Xu, Xu Yang, Yingzhe Peng, Zhiyuan You","submitted_at":"2025-03-10T17:04:14Z","abstract_excerpt":"Enhancing reasoning in Large Multimodal Models (LMMs) faces unique challenges from the complex interplay between visual perception and logical reasoning, particularly in compact 3B-parameter architectures where architectural constraints limit reasoning capacity and modality alignment.\n  While rule-based reinforcement learning (RL) excels in text-only domains, its multimodal extension confronts two critical barriers: (1) data limitations due to ambiguous answers and scarce complex reasoning examples, and (2) degraded foundational reasoning induced by multimodal pretraining. To address these cha"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"text-based reasoning enhancement enables effective multimodal generalization, offering a data-efficient paradigm that bypasses costly high-quality multimodal training data.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"That reasoning skills strengthened via text-only rule-based RL transfer to multimodal inputs without substantial interference from visual perception components or degradation of the foundational reasoning.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"A two-stage RL framework first boosts text reasoning in 3B LMMs then adapts it to multimodal inputs, producing modest benchmark gains of 4.5-4.8%.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"Strengthening reasoning first on text data then transferring to images improves 3B multimodal models without extra multimodal data.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"86c3f6d99d27df1b124bf81641972e85a603eca4101442a220dc3281f3f85af3"},"source":{"id":"2503.07536","kind":"arxiv","version":2},"verdict":{"id":"9c7ebaba-1a3d-46f4-a4f2-4d7f1402af82","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-16T15:11:35.828514Z","strongest_claim":"text-based reasoning enhancement enables effective multimodal generalization, offering a data-efficient paradigm that bypasses costly high-quality multimodal training data.","one_line_summary":"A two-stage RL framework first boosts text reasoning in 3B LMMs then adapts it to multimodal inputs, producing modest benchmark gains of 4.5-4.8%.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"That reasoning skills strengthened via text-only rule-based RL transfer to multimodal inputs without substantial interference from visual perception components or degradation of the foundational reasoning.","pith_extraction_headline":"Strengthening reasoning first on text data then transferring to images improves 3B multimodal models without extra multimodal data."},"references":{"count":117,"sample":[{"doi":"","year":null,"title":"GPT-4 Technical Report","work_id":"b928e041-6991-4c08-8c81-0359e4097c7b","ref_index":1,"cited_arxiv_id":"2303.08774","is_internal_anchor":true},{"doi":"","year":2015,"title":"Lawrence Zitnick, Devi Parikh, and Dhruv Ba- tra","work_id":"7945da46-3915-4108-853a-ab77b7e0cd3c","ref_index":2,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":null,"title":"Flamingo: A visual language model for few-shot learning","work_id":"1863a034-835c-4304-b6c6-32734dcf9e1a","ref_index":3,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2023,"title":"Qwen Technical Report","work_id":"bb1fd52f-6b2f-437c-9516-37bdf6eb9be8","ref_index":4,"cited_arxiv_id":"2309.16609","is_internal_anchor":true},{"doi":"","year":2023,"title":"Qwen-VL: A Versatile Vision-Language Model for Understanding, Localization, Text Reading, and Beyond","work_id":"cbc2bb21-b6bb-46c0-80bf-107e195ffe10","ref_index":5,"cited_arxiv_id":"2308.12966","is_internal_anchor":true}],"resolved_work":117,"snapshot_sha256":"19d1ebaaa5411c244a51a482d94aacad3b045979c3b0e1835cfd65922d00882d","internal_anchors":31},"formal_canon":{"evidence_count":2,"snapshot_sha256":"1f56fb3c6ceb513d5107c257c4adc2804883b4df9c9db298f66864c895bfa6d4"},"author_claims":{"count":0,"strong_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"builder_version":"pith-number-builder-2026-05-17-v1"}