{"paper":{"title":"VL-Rethinker: Incentivizing Self-Reflection of Vision-Language Models with Reinforcement Learning","license":"http://creativecommons.org/licenses/by/4.0/","headline":"Reinforcement learning with selective replay and forced rethinking steps lets vision-language models reflect on their answers and reach new highs on multimodal math benchmarks.","cross_cats":["cs.AI"],"primary_cat":"cs.LG","authors_text":"Chao Qu, Fangzhen Lin, Haozhe Wang, Wei Chu, Wenhu Chen, Zuming Huang","submitted_at":"2025-04-10T17:41:56Z","abstract_excerpt":"Recently, slow-thinking systems like GPT-o1 and DeepSeek-R1 have demonstrated great potential in solving challenging problems through explicit reflection. They significantly outperform the best fast-thinking models, such as GPT-4o, on various math and science benchmarks. However, their multimodal reasoning capabilities remain on par with fast-thinking models. For instance, GPT-o1's performance on benchmarks like MathVista, MathVerse, and MathVision is similar to fast-thinking models. In this paper, we aim to enhance the slow-thinking capabilities of vision-language models using reinforcement l"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"By combining Selective Sample Replay and Forced Rethinking in RL training, VL-Rethinker advances state-of-the-art scores on MathVista to 80.4% and MathVerse to 63.5%, achieving open-source SoTA on MathVision, MMMU-Pro, EMMA, and MEGA-Bench.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"That the reported gains stem primarily from increased self-reflection and slow-thinking rather than from other side effects of the RL setup or from benchmark-specific optimizations.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"VL-Rethinker reaches 80.4% on MathVista and 63.5% on MathVerse by adapting GRPO with Selective Sample Replay and Forced Rethinking to promote self-reflection in vision-language models without distillation.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"Reinforcement learning with selective replay and forced rethinking steps lets vision-language models reflect on their answers and reach new highs on multimodal math benchmarks.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"b0c4414d16041412c92919d0cf6aa837c74158d31ffffaf3905eb35a98691f03"},"source":{"id":"2504.08837","kind":"arxiv","version":3},"verdict":{"id":"fb142a08-24f6-4a45-93ca-6e6124f8f84e","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-15T06:08:53.147112Z","strongest_claim":"By combining Selective Sample Replay and Forced Rethinking in RL training, VL-Rethinker advances state-of-the-art scores on MathVista to 80.4% and MathVerse to 63.5%, achieving open-source SoTA on MathVision, MMMU-Pro, EMMA, and MEGA-Bench.","one_line_summary":"VL-Rethinker reaches 80.4% on MathVista and 63.5% on MathVerse by adapting GRPO with Selective Sample Replay and Forced Rethinking to promote self-reflection in vision-language models without distillation.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"That the reported gains stem primarily from increased self-reflection and slow-thinking rather than from other side effects of the RL setup or from benchmark-specific optimizations.","pith_extraction_headline":"Reinforcement learning with selective replay and forced rethinking steps lets vision-language models reflect on their answers and reach new highs on multimodal math benchmarks."},"references":{"count":0,"sample":[],"resolved_work":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57","internal_anchors":0},"formal_canon":{"evidence_count":2,"snapshot_sha256":"8ceeed42e106089115381c36b37aff4df123d475d9489a7c85c2ad5702295eae"},"author_claims":{"count":0,"strong_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"builder_version":"pith-number-builder-2026-05-17-v1"}