{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2025:LCB3UHAAN6GMCNM5457GPWNF6W","short_pith_number":"pith:LCB3UHAA","schema_version":"1.0","canonical_sha256":"5883ba1c006f8cc1359de77e67d9a5f5b64427899b130d9f6a57ce29a199434e","source":{"kind":"arxiv","id":"2504.08837","version":3},"attestation_state":"computed","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"},"verification_status":{"content_addressed":true,"pith_receipt":true,"author_attested":false,"weak_author_claims":0,"strong_author_claims":0,"externally_anchored":false,"storage_verified":false,"citation_signatures":0,"replication_records":0,"graph_snapshot":true,"references_resolved":false,"formal_links_present":true},"canonical_record":{"source":{"id":"2504.08837","kind":"arxiv","version":3},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.LG","submitted_at":"2025-04-10T17:41:56Z","cross_cats_sorted":["cs.AI"],"title_canon_sha256":"44ab9014d6ac1819ca9747e6da46d768642c25e8940acc1356b7d8608ad8420c","abstract_canon_sha256":"c32ff01e1d574c8407f7d747be89e3134028b09fc84bdb3106f98de9fc65a742"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-05-17T23:38:53.291529Z","signature_b64":"0Ff4uNfcoPDKTyfR1lBguygYU6b7kNM58yY0mEe2JXdKmOJJklcUWxzdRnibryfzJEeeZeftfkMrsKwyVQNiAw==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"5883ba1c006f8cc1359de77e67d9a5f5b64427899b130d9f6a57ce29a199434e","last_reissued_at":"2026-05-17T23:38:53.290891Z","signature_status":"signed_v1","first_computed_at":"2026-05-17T23:38:53.290891Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"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"},"aliases":[{"alias_kind":"arxiv","alias_value":"2504.08837","created_at":"2026-05-17T23:38:53.290980+00:00"},{"alias_kind":"arxiv_version","alias_value":"2504.08837v3","created_at":"2026-05-17T23:38:53.290980+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2504.08837","created_at":"2026-05-17T23:38:53.290980+00:00"},{"alias_kind":"pith_short_12","alias_value":"LCB3UHAAN6GM","created_at":"2026-05-18T12:33:37.589309+00:00"},{"alias_kind":"pith_short_16","alias_value":"LCB3UHAAN6GMCNM5","created_at":"2026-05-18T12:33:37.589309+00:00"},{"alias_kind":"pith_short_8","alias_value":"LCB3UHAA","created_at":"2026-05-18T12:33:37.589309+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":50,"internal_anchor_count":26,"sample":[{"citing_arxiv_id":"2605.21988","citing_title":"Learning Spatiotemporal Sensitivity in Video LLMs via Counterfactual Reinforcement Learning","ref_index":39,"is_internal_anchor":true},{"citing_arxiv_id":"2605.21924","citing_title":"Visual-Advantage On-Policy Distillation for Vision-Language Models","ref_index":7,"is_internal_anchor":true},{"citing_arxiv_id":"2605.15864","citing_title":"Are VLMs Seeing or Just Saying? Uncovering the Illusion of Visual Re-examination","ref_index":15,"is_internal_anchor":true},{"citing_arxiv_id":"2605.15764","citing_title":"GRASP: Learning to Ground Social Reasoning in Multi-Person Non-Verbal Interactions","ref_index":78,"is_internal_anchor":true},{"citing_arxiv_id":"2605.18903","citing_title":"Reasoning Portability: Guiding Continual Learning for MLLMs in the RLVR Era","ref_index":48,"is_internal_anchor":true},{"citing_arxiv_id":"2605.18641","citing_title":"Leveraging Latent Visual Reasoning in Silence","ref_index":31,"is_internal_anchor":true},{"citing_arxiv_id":"2605.12309","citing_title":"G$^2$TR: Generation-Guided Visual Token Reduction for Separate-Encoder Unified Multimodal Models","ref_index":30,"is_internal_anchor":true},{"citing_arxiv_id":"2605.13169","citing_title":"PanoWorld: Towards Spatial Supersensing in 360$^\\circ$ Panorama World","ref_index":40,"is_internal_anchor":true},{"citing_arxiv_id":"2506.11991","citing_title":"VGR: Visual Grounded Reasoning","ref_index":46,"is_internal_anchor":true},{"citing_arxiv_id":"2503.17352","citing_title":"OpenVLThinker: Complex Vision-Language Reasoning via Iterative SFT-RL Cycles","ref_index":72,"is_internal_anchor":true},{"citing_arxiv_id":"2507.06448","citing_title":"Perception-Aware Policy Optimization for Multimodal Reasoning","ref_index":32,"is_internal_anchor":true},{"citing_arxiv_id":"2508.13998","citing_title":"Embodied-R1: Reinforced Embodied Reasoning for General Robotic Manipulation","ref_index":35,"is_internal_anchor":true},{"citing_arxiv_id":"2510.04225","citing_title":"Locate-Then-Examine: Grounded Region Reasoning Improves Detection of AI-Generated Images","ref_index":21,"is_internal_anchor":true},{"citing_arxiv_id":"2511.00710","citing_title":"Does RLVR Extend Reasoning Boundaries? Investigating Capability Expansion in Vision-Language Models","ref_index":6,"is_internal_anchor":true},{"citing_arxiv_id":"2511.13026","citing_title":"REVISOR: Beyond Textual Reflection, Towards Multimodal Introspective Reasoning in Long-Form Video Understanding","ref_index":42,"is_internal_anchor":true},{"citing_arxiv_id":"2511.19972","citing_title":"Boosting Reasoning in Large Multimodal Models via Activation Replay","ref_index":44,"is_internal_anchor":true},{"citing_arxiv_id":"2511.22396","citing_title":"Asking like Socrates: Socrates helps VLMs understand remote sensing images","ref_index":34,"is_internal_anchor":true},{"citing_arxiv_id":"2512.03043","citing_title":"OneThinker: All-in-one Reasoning Model for Image and Video","ref_index":44,"is_internal_anchor":true},{"citing_arxiv_id":"2512.12623","citing_title":"Reasoning Within the Mind: Dynamic Multimodal Interleaving in Latent Space","ref_index":21,"is_internal_anchor":true},{"citing_arxiv_id":"2601.04442","citing_title":"Addressing Overthinking in Large Vision-Language Models via Gated Perception-Reasoning Optimization","ref_index":12,"is_internal_anchor":true},{"citing_arxiv_id":"2601.06803","citing_title":"Forest Before Trees: Latent Superposition for Efficient Visual Reasoning","ref_index":17,"is_internal_anchor":true},{"citing_arxiv_id":"2511.05271","citing_title":"DeepEyesV2: Toward Agentic Multimodal Model","ref_index":45,"is_internal_anchor":true},{"citing_arxiv_id":"2602.09850","citing_title":"Towards Explainable Industrial Anomaly Detection via Knowledge-Guided Latent Reasoning","ref_index":20,"is_internal_anchor":true},{"citing_arxiv_id":"2509.24251","citing_title":"Latent Visual Reasoning","ref_index":23,"is_internal_anchor":true},{"citing_arxiv_id":"2506.07044","citing_title":"Lingshu: A Generalist Foundation Model for Unified Multimodal Medical Understanding and Reasoning","ref_index":10,"is_internal_anchor":true}]},"formal_canon":{"evidence_count":2,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/LCB3UHAAN6GMCNM5457GPWNF6W","json":"https://pith.science/pith/LCB3UHAAN6GMCNM5457GPWNF6W.json","graph_json":"https://pith.science/api/pith-number/LCB3UHAAN6GMCNM5457GPWNF6W/graph.json","events_json":"https://pith.science/api/pith-number/LCB3UHAAN6GMCNM5457GPWNF6W/events.json","paper":"https://pith.science/paper/LCB3UHAA"},"agent_actions":{"view_html":"https://pith.science/pith/LCB3UHAAN6GMCNM5457GPWNF6W","download_json":"https://pith.science/pith/LCB3UHAAN6GMCNM5457GPWNF6W.json","view_paper":"https://pith.science/paper/LCB3UHAA","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2504.08837&json=true","fetch_graph":"https://pith.science/api/pith-number/LCB3UHAAN6GMCNM5457GPWNF6W/graph.json","fetch_events":"https://pith.science/api/pith-number/LCB3UHAAN6GMCNM5457GPWNF6W/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/LCB3UHAAN6GMCNM5457GPWNF6W/action/timestamp_anchor","attest_storage":"https://pith.science/pith/LCB3UHAAN6GMCNM5457GPWNF6W/action/storage_attestation","attest_author":"https://pith.science/pith/LCB3UHAAN6GMCNM5457GPWNF6W/action/author_attestation","sign_citation":"https://pith.science/pith/LCB3UHAAN6GMCNM5457GPWNF6W/action/citation_signature","submit_replication":"https://pith.science/pith/LCB3UHAAN6GMCNM5457GPWNF6W/action/replication_record"}},"created_at":"2026-05-17T23:38:53.290980+00:00","updated_at":"2026-05-17T23:38:53.290980+00:00"}