{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2025:R2U4FDUROUVZCFF5BNEEKHYNLN","short_pith_number":"pith:R2U4FDUR","schema_version":"1.0","canonical_sha256":"8ea9c28e91752b9114bd0b48451f0d5b639ff2abc94e924e5782d60be48f645d","source":{"kind":"arxiv","id":"2503.23383","version":1},"attestation_state":"computed","paper":{"title":"ToRL: Scaling Tool-Integrated RL","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":[],"primary_cat":"cs.CL","authors_text":"Haoyang Zou, Pengfei Liu, Xuefeng Li","submitted_at":"2025-03-30T10:16:25Z","abstract_excerpt":"We introduce ToRL (Tool-Integrated Reinforcement Learning), a framework for training large language models (LLMs) to autonomously use computational tools via reinforcement learning. Unlike supervised fine-tuning, ToRL allows models to explore and discover optimal strategies for tool use. Experiments with Qwen2.5-Math models show significant improvements: ToRL-7B reaches 43.3\\% accuracy on AIME~24, surpassing reinforcement learning without tool integration by 14\\% and the best existing Tool-Integrated Reasoning (TIR) model by 17\\%. Further analysis reveals emergent behaviors such as strategic t"},"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":false},"canonical_record":{"source":{"id":"2503.23383","kind":"arxiv","version":1},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.CL","submitted_at":"2025-03-30T10:16:25Z","cross_cats_sorted":[],"title_canon_sha256":"e8354db1f4e53bc9a6263fb6951d9dfb53e140b38e81e7487764f7c49b371ca9","abstract_canon_sha256":"e665d2b221fdc9ae39d0eeeb54205e96a3a141fe2bf6efc054967d324adef4ae"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T10:41:35.335932Z","signature_b64":"vsPLGK6OtzKk9pDJ/XDbkQjPjc0P2y8mDPbEqT6tC2/TuCZ5Yl32YGhxMRdLa/g+qhiX5DzGzhnLnyKnzFpIAw==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"8ea9c28e91752b9114bd0b48451f0d5b639ff2abc94e924e5782d60be48f645d","last_reissued_at":"2026-07-05T10:41:35.335436Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T10:41:35.335436Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"ToRL: Scaling Tool-Integrated RL","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":[],"primary_cat":"cs.CL","authors_text":"Haoyang Zou, Pengfei Liu, Xuefeng Li","submitted_at":"2025-03-30T10:16:25Z","abstract_excerpt":"We introduce ToRL (Tool-Integrated Reinforcement Learning), a framework for training large language models (LLMs) to autonomously use computational tools via reinforcement learning. Unlike supervised fine-tuning, ToRL allows models to explore and discover optimal strategies for tool use. Experiments with Qwen2.5-Math models show significant improvements: ToRL-7B reaches 43.3\\% accuracy on AIME~24, surpassing reinforcement learning without tool integration by 14\\% and the best existing Tool-Integrated Reasoning (TIR) model by 17\\%. Further analysis reveals emergent behaviors such as strategic t"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2503.23383","kind":"arxiv","version":1},"verdict":{"id":null,"model_set":{},"created_at":null,"strongest_claim":"","one_line_summary":"","pipeline_version":null,"weakest_assumption":"","pith_extraction_headline":""},"integrity":{"clean":true,"summary":{"advisory":0,"critical":0,"by_detector":{},"informational":0},"endpoint":"/pith/2503.23383/integrity.json","findings":[],"available":true,"detectors_run":[],"snapshot_sha256":"c28c3603d3b5d939e8dc4c7e95fa8dfce3d595e45f758748cecf8e644a296938"},"references":{"count":0,"sample":[],"resolved_work":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57","internal_anchors":0},"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"},"aliases":[{"alias_kind":"arxiv","alias_value":"2503.23383","created_at":"2026-07-05T10:41:35.335499+00:00"},{"alias_kind":"arxiv_version","alias_value":"2503.23383v1","created_at":"2026-07-05T10:41:35.335499+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2503.23383","created_at":"2026-07-05T10:41:35.335499+00:00"},{"alias_kind":"pith_short_12","alias_value":"R2U4FDUROUVZ","created_at":"2026-07-05T10:41:35.335499+00:00"},{"alias_kind":"pith_short_16","alias_value":"R2U4FDUROUVZCFF5","created_at":"2026-07-05T10:41:35.335499+00:00"},{"alias_kind":"pith_short_8","alias_value":"R2U4FDUR","created_at":"2026-07-05T10:41:35.335499+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":32,"internal_anchor_count":0,"sample":[{"citing_arxiv_id":"2606.26027","citing_title":"Why Multi-Step Tool-Use Reinforcement Learning Collapses and How Supervisory Signals Fix It","ref_index":2,"is_internal_anchor":false},{"citing_arxiv_id":"2606.21943","citing_title":"Modularized Reinforcement Learning on LLMs: From MDP Creation to Exploration and Learning","ref_index":107,"is_internal_anchor":false},{"citing_arxiv_id":"2606.12908","citing_title":"SENTINEL: Failure-Driven Reinforcement Learning for Training Tool-Using Language Model Agents","ref_index":52,"is_internal_anchor":false},{"citing_arxiv_id":"2606.12384","citing_title":"APPO: Agentic Procedural Policy Optimization","ref_index":41,"is_internal_anchor":false},{"citing_arxiv_id":"2606.11652","citing_title":"IAPO: Input Attribution-Aware Policy Optimization for Tool Use in Small Multimodal Agents","ref_index":38,"is_internal_anchor":false},{"citing_arxiv_id":"2606.10875","citing_title":"Pushing the Limits of LLM Tool Calling via Experiential Knowledge Integration and Activation","ref_index":2,"is_internal_anchor":false},{"citing_arxiv_id":"2606.09371","citing_title":"Capability-Aligned Hierarchical Learning for Tool-Augmented LLMs","ref_index":13,"is_internal_anchor":false},{"citing_arxiv_id":"2606.07074","citing_title":"SlimSearcher: Training Efficiency-Aware Web Agents via Adaptive Reward Gating","ref_index":6,"is_internal_anchor":false},{"citing_arxiv_id":"2607.01084","citing_title":"Can Agents Generalize to the Open World? Unveiling the Fragility of Static Training in Tool Use","ref_index":53,"is_internal_anchor":false},{"citing_arxiv_id":"2606.03762","citing_title":"Tool-Aware Optimization with Entropy Guidance for Efficient Agentic Reinforcement Learning","ref_index":4,"is_internal_anchor":false},{"citing_arxiv_id":"2606.31650","citing_title":"ECHO: Prune To Act, Trace To Learn With Selective Turn Memory In Agentic RL","ref_index":11,"is_internal_anchor":false},{"citing_arxiv_id":"2605.27788","citing_title":"Knowing When to Ask: Segment-Level Credit Assignment for LLM Tool Use","ref_index":7,"is_internal_anchor":false},{"citing_arxiv_id":"2605.28774","citing_title":"Agent Explorative Policy Optimization for Multimodal Agentic Reasoning","ref_index":64,"is_internal_anchor":false},{"citing_arxiv_id":"2605.28774","citing_title":"Agent Explorative Policy Optimization for Multimodal Agentic Reasoning","ref_index":65,"is_internal_anchor":false},{"citing_arxiv_id":"2605.10787","citing_title":"ComplexMCP: Evaluation of LLM Agents in Dynamic, Interdependent, and Large-Scale Tool Sandbox","ref_index":4,"is_internal_anchor":false},{"citing_arxiv_id":"2605.17734","citing_title":"Harnessing LLM Agents with Skill Programs","ref_index":17,"is_internal_anchor":false},{"citing_arxiv_id":"2509.02547","citing_title":"The Landscape of Agentic Reinforcement Learning for LLMs: A Survey","ref_index":109,"is_internal_anchor":false},{"citing_arxiv_id":"2509.18847","citing_title":"Failure Makes the Agent Stronger: Enhancing Accuracy through Structured Reflection for Reliable Tool Interactions","ref_index":12,"is_internal_anchor":false},{"citing_arxiv_id":"2510.22977","citing_title":"The Reasoning Trap: How Enhancing LLM Reasoning Amplifies Tool Hallucination","ref_index":7,"is_internal_anchor":false},{"citing_arxiv_id":"2509.08827","citing_title":"A Survey of Reinforcement Learning for Large Reasoning Models","ref_index":290,"is_internal_anchor":false},{"citing_arxiv_id":"2504.21776","citing_title":"WebThinker: Empowering Large Reasoning Models with Deep Research Capability","ref_index":27,"is_internal_anchor":false},{"citing_arxiv_id":"2603.16876","citing_title":"Multi-Modal Multi-Agent Reinforcement Learning for Radiology Report Generation","ref_index":60,"is_internal_anchor":false},{"citing_arxiv_id":"2604.02794","citing_title":"CharTool: Tool-Integrated Visual Reasoning for Chart Understanding","ref_index":25,"is_internal_anchor":false},{"citing_arxiv_id":"2504.11536","citing_title":"ReTool: Reinforcement Learning for Strategic Tool Use in LLMs","ref_index":8,"is_internal_anchor":false},{"citing_arxiv_id":"2509.02544","citing_title":"UI-TARS-2 Technical Report: Advancing GUI Agent with Multi-Turn Reinforcement Learning","ref_index":34,"is_internal_anchor":false}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/R2U4FDUROUVZCFF5BNEEKHYNLN","json":"https://pith.science/pith/R2U4FDUROUVZCFF5BNEEKHYNLN.json","graph_json":"https://pith.science/api/pith-number/R2U4FDUROUVZCFF5BNEEKHYNLN/graph.json","events_json":"https://pith.science/api/pith-number/R2U4FDUROUVZCFF5BNEEKHYNLN/events.json","paper":"https://pith.science/paper/R2U4FDUR"},"agent_actions":{"view_html":"https://pith.science/pith/R2U4FDUROUVZCFF5BNEEKHYNLN","download_json":"https://pith.science/pith/R2U4FDUROUVZCFF5BNEEKHYNLN.json","view_paper":"https://pith.science/paper/R2U4FDUR","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2503.23383&json=true","fetch_graph":"https://pith.science/api/pith-number/R2U4FDUROUVZCFF5BNEEKHYNLN/graph.json","fetch_events":"https://pith.science/api/pith-number/R2U4FDUROUVZCFF5BNEEKHYNLN/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/R2U4FDUROUVZCFF5BNEEKHYNLN/action/timestamp_anchor","attest_storage":"https://pith.science/pith/R2U4FDUROUVZCFF5BNEEKHYNLN/action/storage_attestation","attest_author":"https://pith.science/pith/R2U4FDUROUVZCFF5BNEEKHYNLN/action/author_attestation","sign_citation":"https://pith.science/pith/R2U4FDUROUVZCFF5BNEEKHYNLN/action/citation_signature","submit_replication":"https://pith.science/pith/R2U4FDUROUVZCFF5BNEEKHYNLN/action/replication_record"}},"created_at":"2026-07-05T10:41:35.335499+00:00","updated_at":"2026-07-05T10:41:35.335499+00:00"}