{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2025:K6JGKNLIUJKGSEVPTRP3LSOXJG","short_pith_number":"pith:K6JGKNLI","schema_version":"1.0","canonical_sha256":"5792653568a2546912af9c5fb5c9d749b47e3aac31f3591d0f53730ce5221e15","source":{"kind":"arxiv","id":"2505.08617","version":2},"attestation_state":"computed","paper":{"title":"OpenThinkIMG: Learning to Think with Images via Visual Tool Reinforcement Learning","license":"http://creativecommons.org/licenses/by-sa/4.0/","headline":"Reinforcement learning on visual tool feedback lets a small LVLM learn adaptive tool-use policies that outperform supervised training and some larger models on chart reasoning.","cross_cats":[],"primary_cat":"cs.CV","authors_text":"Guanjie Chen, Jiawei Gu, Juntao Li, Jun Zhang, Linjie Li, Mingyang Song, Xiaoye Qu, Yu Cheng, Yunzhuo Hao, Zhaochen Su, Zhengyuan Yang","submitted_at":"2025-05-13T14:35:51Z","abstract_excerpt":"While humans can flexibly leverage interactive visual cognition for complex problem-solving, enabling Large Vision-Language Models (LVLMs) to learn similarly adaptive behaviors with visual tools remains challenging. A significant hurdle is the current lack of standardized infrastructure, which hinders integrating diverse tools, generating rich interaction data, and training robust agents effectively. To address these gaps, we introduce OpenThinkIMG, the first open-source, comprehensive end-to-end framework for tool-augmented LVLMs. It features standardized vision tool interfaces, scalable traj"},"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":"2505.08617","kind":"arxiv","version":2},"metadata":{"license":"http://creativecommons.org/licenses/by-sa/4.0/","primary_cat":"cs.CV","submitted_at":"2025-05-13T14:35:51Z","cross_cats_sorted":[],"title_canon_sha256":"3ba2bef8bf3e1210a9ca37ae961fff081ff482833b3e6e2b4b0ceb9e86148ac2","abstract_canon_sha256":"d3cec06cfc35f7bb27006d3c59e03cf69c4435b409847f692a8c93cf6d4e4e2c"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-05-17T23:38:46.421874Z","signature_b64":"TCIBAnFdFihd2r5nBXntN5TrgxsDf7kPwgfz8i4MBR/IPQTJOchPEfyEh4i0x9oUouyctiN0HwWvYyADHgP0CA==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"5792653568a2546912af9c5fb5c9d749b47e3aac31f3591d0f53730ce5221e15","last_reissued_at":"2026-05-17T23:38:46.421271Z","signature_status":"signed_v1","first_computed_at":"2026-05-17T23:38:46.421271Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"OpenThinkIMG: Learning to Think with Images via Visual Tool Reinforcement Learning","license":"http://creativecommons.org/licenses/by-sa/4.0/","headline":"Reinforcement learning on visual tool feedback lets a small LVLM learn adaptive tool-use policies that outperform supervised training and some larger models on chart reasoning.","cross_cats":[],"primary_cat":"cs.CV","authors_text":"Guanjie Chen, Jiawei Gu, Juntao Li, Jun Zhang, Linjie Li, Mingyang Song, Xiaoye Qu, Yu Cheng, Yunzhuo Hao, Zhaochen Su, Zhengyuan Yang","submitted_at":"2025-05-13T14:35:51Z","abstract_excerpt":"While humans can flexibly leverage interactive visual cognition for complex problem-solving, enabling Large Vision-Language Models (LVLMs) to learn similarly adaptive behaviors with visual tools remains challenging. A significant hurdle is the current lack of standardized infrastructure, which hinders integrating diverse tools, generating rich interaction data, and training robust agents effectively. To address these gaps, we introduce OpenThinkIMG, the first open-source, comprehensive end-to-end framework for tool-augmented LVLMs. It features standardized vision tool interfaces, scalable traj"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"Our RL-trained agent, built upon a Qwen2-VL-2B, significantly outperforms its SFT-initialized counterpart (+28.83 points) and surpasses established supervised tool-learning baselines like Taco and CogCom by an average of +12.7 points. Notably, it also surpasses prominent closed-source models like GPT-4.1 by +8.68 accuracy points.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"The assumption that feedback from tool interactions on chart reasoning tasks will produce policies that generalize to other visual domains and tool sets without additional tuning or domain-specific reward shaping.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"OpenThinkIMG and V-ToolRL enable LVLMs to learn adaptive visual tool use via RL, yielding a Qwen2-VL-2B agent that beats its SFT version by 28.83 points and GPT-4.1 by 8.68 points on chart reasoning.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"Reinforcement learning on visual tool feedback lets a small LVLM learn adaptive tool-use policies that outperform supervised training and some larger models on chart reasoning.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"fb754c29782ff7029a4d9d066f2970e93b83b2eac7cb39a55770a9f0290492ca"},"source":{"id":"2505.08617","kind":"arxiv","version":2},"verdict":{"id":"4d787b5f-af64-445b-b60a-144d46e6c0b1","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-16T22:07:29.872962Z","strongest_claim":"Our RL-trained agent, built upon a Qwen2-VL-2B, significantly outperforms its SFT-initialized counterpart (+28.83 points) and surpasses established supervised tool-learning baselines like Taco and CogCom by an average of +12.7 points. Notably, it also surpasses prominent closed-source models like GPT-4.1 by +8.68 accuracy points.","one_line_summary":"OpenThinkIMG and V-ToolRL enable LVLMs to learn adaptive visual tool use via RL, yielding a Qwen2-VL-2B agent that beats its SFT version by 28.83 points and GPT-4.1 by 8.68 points on chart reasoning.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"The assumption that feedback from tool interactions on chart reasoning tasks will produce policies that generalize to other visual domains and tool sets without additional tuning or domain-specific reward shaping.","pith_extraction_headline":"Reinforcement learning on visual tool feedback lets a small LVLM learn adaptive tool-use policies that outperform supervised training and some larger models on chart reasoning."},"references":{"count":0,"sample":[],"resolved_work":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57","internal_anchors":0},"formal_canon":{"evidence_count":2,"snapshot_sha256":"f7afe6eec8f19d01b457ed1e666d8f8853c913b0fa330dfc1e156fa0076274aa"},"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":"2505.08617","created_at":"2026-05-17T23:38:46.421370+00:00"},{"alias_kind":"arxiv_version","alias_value":"2505.08617v2","created_at":"2026-05-17T23:38:46.421370+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2505.08617","created_at":"2026-05-17T23:38:46.421370+00:00"},{"alias_kind":"pith_short_12","alias_value":"K6JGKNLIUJKG","created_at":"2026-05-18T12:33:37.589309+00:00"},{"alias_kind":"pith_short_16","alias_value":"K6JGKNLIUJKGSEVP","created_at":"2026-05-18T12:33:37.589309+00:00"},{"alias_kind":"pith_short_8","alias_value":"K6JGKNLI","created_at":"2026-05-18T12:33:37.589309+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":30,"internal_anchor_count":30,"sample":[{"citing_arxiv_id":"2512.15693","citing_title":"Skyra: AI-Generated Video Detection via Grounded Artifact Reasoning","ref_index":61,"is_internal_anchor":true},{"citing_arxiv_id":"2605.16416","citing_title":"CAVE: A Structured Credit Assignment Approach for Fragmented Visual Evidence Reasoning","ref_index":10,"is_internal_anchor":true},{"citing_arxiv_id":"2605.15792","citing_title":"Reversing the Flow: Generation-to-Understanding Synergy in Large Multimodal Models","ref_index":35,"is_internal_anchor":true},{"citing_arxiv_id":"2605.19528","citing_title":"Towards Camera-Robust 3D Localization: Equation-Anchored Tool-Use for MLLMs","ref_index":33,"is_internal_anchor":true},{"citing_arxiv_id":"2605.19538","citing_title":"CaptchaMind: Training CAPTCHA Solvers via Reinforcement Learning with Explicit Reasoning Supervision","ref_index":9,"is_internal_anchor":true},{"citing_arxiv_id":"2605.19852","citing_title":"Are Tools Always Beneficial? Learning to Invoke Tools Adaptively for Dual-Mode Multimodal LLM Reasoning","ref_index":38,"is_internal_anchor":true},{"citing_arxiv_id":"2509.02547","citing_title":"The Landscape of Agentic Reinforcement Learning for LLMs: A Survey","ref_index":254,"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":20,"is_internal_anchor":true},{"citing_arxiv_id":"2512.08980","citing_title":"Training Multi-Image Vision Agents via End2End Reinforcement Learning","ref_index":36,"is_internal_anchor":true},{"citing_arxiv_id":"2512.12623","citing_title":"Reasoning Within the Mind: Dynamic Multimodal Interleaving in Latent Space","ref_index":26,"is_internal_anchor":true},{"citing_arxiv_id":"2512.16300","citing_title":"Code-in-the-Loop Forensics: Agentic Tool Use for Image Forgery Detection","ref_index":35,"is_internal_anchor":true},{"citing_arxiv_id":"2512.16918","citing_title":"AdaTooler-V: Adaptive Tool-Use for Images and Videos","ref_index":55,"is_internal_anchor":true},{"citing_arxiv_id":"2601.15724","citing_title":"VideoThinker: Building Agentic VideoLLMs with LLM-Guided Tool Reasoning","ref_index":27,"is_internal_anchor":true},{"citing_arxiv_id":"2508.05748","citing_title":"WebWatcher: Breaking New Frontier of Vision-Language Deep Research Agent","ref_index":23,"is_internal_anchor":true},{"citing_arxiv_id":"2605.12163","citing_title":"Self-Consistent Latent Reasoning: Long Latent Sequence Reasoning for Vision-Language Model","ref_index":40,"is_internal_anchor":true},{"citing_arxiv_id":"2604.24919","citing_title":"Agentic AI for Remote Sensing: Technical Challenges and Research Directions","ref_index":107,"is_internal_anchor":true},{"citing_arxiv_id":"2604.03307","citing_title":"V-Reflection: Transforming MLLMs from Passive Observers to Active Interrogators","ref_index":20,"is_internal_anchor":true},{"citing_arxiv_id":"2605.12163","citing_title":"Self-Consistent Latent Reasoning: Long Latent Sequence Reasoning for Vision-Language Model","ref_index":40,"is_internal_anchor":true},{"citing_arxiv_id":"2605.08735","citing_title":"CollabVR: Collaborative Video Reasoning with Vision-Language and Video Generation Models","ref_index":29,"is_internal_anchor":true},{"citing_arxiv_id":"2604.24919","citing_title":"Agentic AI for Remote Sensing: Technical Challenges and Research Directions","ref_index":107,"is_internal_anchor":true},{"citing_arxiv_id":"2604.24339","citing_title":"See Further, Think Deeper: Advancing VLM's Reasoning Ability with Low-level Visual Cues and Reflection","ref_index":44,"is_internal_anchor":true},{"citing_arxiv_id":"2604.22875","citing_title":"SketchVLM: Vision language models can annotate images to explain thoughts and guide users","ref_index":40,"is_internal_anchor":true},{"citing_arxiv_id":"2604.19945","citing_title":"Visual Reasoning through Tool-supervised Reinforcement Learning","ref_index":23,"is_internal_anchor":true},{"citing_arxiv_id":"2604.19632","citing_title":"CreatiParser: Generative Image Parsing of Raster Graphic Designs into Editable Layers","ref_index":38,"is_internal_anchor":true},{"citing_arxiv_id":"2604.07518","citing_title":"Decompose, Look, and Reason: Reinforced Latent Reasoning for VLMs","ref_index":8,"is_internal_anchor":true}]},"formal_canon":{"evidence_count":2,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/K6JGKNLIUJKGSEVPTRP3LSOXJG","json":"https://pith.science/pith/K6JGKNLIUJKGSEVPTRP3LSOXJG.json","graph_json":"https://pith.science/api/pith-number/K6JGKNLIUJKGSEVPTRP3LSOXJG/graph.json","events_json":"https://pith.science/api/pith-number/K6JGKNLIUJKGSEVPTRP3LSOXJG/events.json","paper":"https://pith.science/paper/K6JGKNLI"},"agent_actions":{"view_html":"https://pith.science/pith/K6JGKNLIUJKGSEVPTRP3LSOXJG","download_json":"https://pith.science/pith/K6JGKNLIUJKGSEVPTRP3LSOXJG.json","view_paper":"https://pith.science/paper/K6JGKNLI","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2505.08617&json=true","fetch_graph":"https://pith.science/api/pith-number/K6JGKNLIUJKGSEVPTRP3LSOXJG/graph.json","fetch_events":"https://pith.science/api/pith-number/K6JGKNLIUJKGSEVPTRP3LSOXJG/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/K6JGKNLIUJKGSEVPTRP3LSOXJG/action/timestamp_anchor","attest_storage":"https://pith.science/pith/K6JGKNLIUJKGSEVPTRP3LSOXJG/action/storage_attestation","attest_author":"https://pith.science/pith/K6JGKNLIUJKGSEVPTRP3LSOXJG/action/author_attestation","sign_citation":"https://pith.science/pith/K6JGKNLIUJKGSEVPTRP3LSOXJG/action/citation_signature","submit_replication":"https://pith.science/pith/K6JGKNLIUJKGSEVPTRP3LSOXJG/action/replication_record"}},"created_at":"2026-05-17T23:38:46.421370+00:00","updated_at":"2026-05-17T23:38:46.421370+00:00"}