{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2025:76D4Y36TSRHSVXGK45AWLZT6RH","short_pith_number":"pith:76D4Y36T","schema_version":"1.0","canonical_sha256":"ff87cc6fd3944f2adccae74165e67e89c8d0d17fadcdbdb40fa4b93da5b0b0b1","source":{"kind":"arxiv","id":"2509.02479","version":2},"attestation_state":"computed","paper":{"title":"SimpleTIR: End-to-End Reinforcement Learning for Multi-Turn Tool-Integrated Reasoning","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":[],"primary_cat":"cs.LG","authors_text":"Bo An, Longtao Zheng, Qian Liu, Xiaosen Zheng, Yingru Li, Zejun Ma, Zhenghai Xue","submitted_at":"2025-09-02T16:30:19Z","abstract_excerpt":"Large Language Models (LLMs) can significantly improve their reasoning capabilities by interacting with external tools, a paradigm known as Tool-Integrated Reasoning (TIR). However, extending TIR to multi-turn scenarios using Reinforcement Learning (RL) is often hindered by training instability and performance collapse. We identify that such instability is primarily caused by a distributional drift from external tool feedback, leading to the generation of low-probability tokens. This issue compounds over successive turns, causing catastrophic gradient norm explosions that derail the training p"},"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":"2509.02479","kind":"arxiv","version":2},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.LG","submitted_at":"2025-09-02T16:30:19Z","cross_cats_sorted":[],"title_canon_sha256":"ae078dee364b1fe5910c2b671736a1ee7474488a6f60a470b4b97ef47ebb2eb5","abstract_canon_sha256":"7094a3651824286953d6a8e786ac0de8c290a087ccbc8014449c068abacaac62"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T12:04:19.345759Z","signature_b64":"toxK2Pdne3duk5FrRj+n5xVYGtECN4bRIVISePgp7g9+u63Y+OBQKOBQgjQADe6hYDuST0Z6ryg91UhNZV9DCQ==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"ff87cc6fd3944f2adccae74165e67e89c8d0d17fadcdbdb40fa4b93da5b0b0b1","last_reissued_at":"2026-07-05T12:04:19.345232Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T12:04:19.345232Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"SimpleTIR: End-to-End Reinforcement Learning for Multi-Turn Tool-Integrated Reasoning","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":[],"primary_cat":"cs.LG","authors_text":"Bo An, Longtao Zheng, Qian Liu, Xiaosen Zheng, Yingru Li, Zejun Ma, Zhenghai Xue","submitted_at":"2025-09-02T16:30:19Z","abstract_excerpt":"Large Language Models (LLMs) can significantly improve their reasoning capabilities by interacting with external tools, a paradigm known as Tool-Integrated Reasoning (TIR). However, extending TIR to multi-turn scenarios using Reinforcement Learning (RL) is often hindered by training instability and performance collapse. We identify that such instability is primarily caused by a distributional drift from external tool feedback, leading to the generation of low-probability tokens. This issue compounds over successive turns, causing catastrophic gradient norm explosions that derail the training p"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2509.02479","kind":"arxiv","version":2},"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/2509.02479/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":"2509.02479","created_at":"2026-07-05T12:04:19.345294+00:00"},{"alias_kind":"arxiv_version","alias_value":"2509.02479v2","created_at":"2026-07-05T12:04:19.345294+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2509.02479","created_at":"2026-07-05T12:04:19.345294+00:00"},{"alias_kind":"pith_short_12","alias_value":"76D4Y36TSRHS","created_at":"2026-07-05T12:04:19.345294+00:00"},{"alias_kind":"pith_short_16","alias_value":"76D4Y36TSRHSVXGK","created_at":"2026-07-05T12:04:19.345294+00:00"},{"alias_kind":"pith_short_8","alias_value":"76D4Y36T","created_at":"2026-07-05T12:04:19.345294+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":27,"internal_anchor_count":1,"sample":[{"citing_arxiv_id":"2607.07050","citing_title":"When Top-K Misses the Decision: Tool-Call Drift in Multi-Teacher On-Policy Distillation","ref_index":19,"is_internal_anchor":true},{"citing_arxiv_id":"2606.26027","citing_title":"Why Multi-Step Tool-Use Reinforcement Learning Collapses and How Supervisory Signals Fix It","ref_index":5,"is_internal_anchor":false},{"citing_arxiv_id":"2606.24233","citing_title":"Latent Visual States for Efficient Multimodal Reasoning","ref_index":40,"is_internal_anchor":false},{"citing_arxiv_id":"2606.23678","citing_title":"AIR: Adaptive Interleaved Reasoning with Code in MLLMs","ref_index":8,"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":3,"is_internal_anchor":false},{"citing_arxiv_id":"2606.03382","citing_title":"Local Guidance, Global Impact: Gaussian-Reshaped Trust Region Unlocks Behavior Transitions","ref_index":38,"is_internal_anchor":false},{"citing_arxiv_id":"2606.02132","citing_title":"Learning When Not to Act: Mitigating Tool Abuse in Agentic Reinforcement Learning","ref_index":3,"is_internal_anchor":false},{"citing_arxiv_id":"2605.28751","citing_title":"Extrapolative Weight Averaging Reveals Correctness-Efficiency Frontiers in Code RL","ref_index":45,"is_internal_anchor":false},{"citing_arxiv_id":"2605.22138","citing_title":"Efficient Agentic Reasoning Through Self-Regulated Simulative Planning","ref_index":103,"is_internal_anchor":false},{"citing_arxiv_id":"2601.06943","citing_title":"Watching, Reasoning, and Searching: A Video Deep Research Benchmark on Open Web for Agentic Video Reasoning","ref_index":15,"is_internal_anchor":false},{"citing_arxiv_id":"2601.21468","citing_title":"MemOCR: Layout-Aware Visual Memory for Efficient Long-Horizon Reasoning","ref_index":28,"is_internal_anchor":false},{"citing_arxiv_id":"2605.17734","citing_title":"Harnessing LLM Agents with Skill Programs","ref_index":37,"is_internal_anchor":false},{"citing_arxiv_id":"2509.07969","citing_title":"Mini-o3: Scaling Up Reasoning Patterns and Interaction Turns for Visual Search","ref_index":42,"is_internal_anchor":false},{"citing_arxiv_id":"2512.08980","citing_title":"Training Multi-Image Vision Agents via End2End Reinforcement Learning","ref_index":53,"is_internal_anchor":false},{"citing_arxiv_id":"2601.03555","citing_title":"SCRIBE: Structured Mid-Level Supervision for Tool-Using Language Models","ref_index":6,"is_internal_anchor":false},{"citing_arxiv_id":"2605.08666","citing_title":"The Cancellation Hypothesis in Critic-Free RL: From Outcome Rewards to Token Credits","ref_index":26,"is_internal_anchor":false},{"citing_arxiv_id":"2605.09931","citing_title":"PruneTIR: Inference-Time Tool Call Pruning for Effective yet Efficient Tool-Integrated Reasoning","ref_index":29,"is_internal_anchor":false},{"citing_arxiv_id":"2605.10754","citing_title":"The Agent Use of Agent Beings: Agent Cybernetics Is the Missing Science of Foundation Agents","ref_index":43,"is_internal_anchor":false},{"citing_arxiv_id":"2605.09544","citing_title":"TIDE-Bench: Task-Aware and Diagnostic Evaluation of Tool-Integrated Reasoning","ref_index":13,"is_internal_anchor":false},{"citing_arxiv_id":"2605.06326","citing_title":"Teaching Thinking Models to Reason with Tools: A Full-Pipeline Recipe for Tool-Integrated Reasoning","ref_index":40,"is_internal_anchor":false},{"citing_arxiv_id":"2604.09455","citing_title":"E3-TIR: Enhanced Experience Exploitation for Tool-Integrated Reasoning","ref_index":32,"is_internal_anchor":false},{"citing_arxiv_id":"2605.07331","citing_title":"Rethinking Importance Sampling in LLM Policy Optimization: A Cumulative Token Perspective","ref_index":46,"is_internal_anchor":false},{"citing_arxiv_id":"2605.07725","citing_title":"SOD: Step-wise On-policy Distillation for Small Language Model Agents","ref_index":15,"is_internal_anchor":false},{"citing_arxiv_id":"2604.06777","citing_title":"Walk the Talk: Bridging the Reasoning-Action Gap for Thinking with Images via Multimodal Agentic Policy Optimization","ref_index":52,"is_internal_anchor":false},{"citing_arxiv_id":"2604.08281","citing_title":"When to Trust Tools? Adaptive Tool Trust Calibration For Tool-Integrated Math Reasoning","ref_index":25,"is_internal_anchor":false}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/76D4Y36TSRHSVXGK45AWLZT6RH","json":"https://pith.science/pith/76D4Y36TSRHSVXGK45AWLZT6RH.json","graph_json":"https://pith.science/api/pith-number/76D4Y36TSRHSVXGK45AWLZT6RH/graph.json","events_json":"https://pith.science/api/pith-number/76D4Y36TSRHSVXGK45AWLZT6RH/events.json","paper":"https://pith.science/paper/76D4Y36T"},"agent_actions":{"view_html":"https://pith.science/pith/76D4Y36TSRHSVXGK45AWLZT6RH","download_json":"https://pith.science/pith/76D4Y36TSRHSVXGK45AWLZT6RH.json","view_paper":"https://pith.science/paper/76D4Y36T","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2509.02479&json=true","fetch_graph":"https://pith.science/api/pith-number/76D4Y36TSRHSVXGK45AWLZT6RH/graph.json","fetch_events":"https://pith.science/api/pith-number/76D4Y36TSRHSVXGK45AWLZT6RH/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/76D4Y36TSRHSVXGK45AWLZT6RH/action/timestamp_anchor","attest_storage":"https://pith.science/pith/76D4Y36TSRHSVXGK45AWLZT6RH/action/storage_attestation","attest_author":"https://pith.science/pith/76D4Y36TSRHSVXGK45AWLZT6RH/action/author_attestation","sign_citation":"https://pith.science/pith/76D4Y36TSRHSVXGK45AWLZT6RH/action/citation_signature","submit_replication":"https://pith.science/pith/76D4Y36TSRHSVXGK45AWLZT6RH/action/replication_record"}},"created_at":"2026-07-05T12:04:19.345294+00:00","updated_at":"2026-07-05T12:04:19.345294+00:00"}