{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2025:A6DFUSY3HLG3BGNZ6QE4MI3NK3","short_pith_number":"pith:A6DFUSY3","schema_version":"1.0","canonical_sha256":"07865a4b1b3acdb099b9f409c6236d56ed6c48695edfebc04f96dfcf19143174","source":{"kind":"arxiv","id":"2506.04821","version":1},"attestation_state":"computed","paper":{"title":"LogicPuzzleRL: Cultivating Robust Mathematical Reasoning in LLMs via Reinforcement Learning","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":[],"primary_cat":"cs.LG","authors_text":"Bin Cui, Chengyu Shen, Hao Liang, Hejun Dong, Jingwen Deng, Qijie You, Runming He, Wentao Zhang, Zhen Hao Wong, Zirong Chen","submitted_at":"2025-06-05T09:40:47Z","abstract_excerpt":"Large language models (LLMs) excel at many supervised tasks but often struggle with structured reasoning in unfamiliar settings. This discrepancy suggests that standard fine-tuning pipelines may instill narrow, domain-specific heuristics rather than fostering general-purpose thinking strategies. In this work, we propose a \"play to learn\" framework that fine-tunes LLMs through reinforcement learning on a suite of seven custom logic puzzles, each designed to cultivate distinct reasoning skills such as constraint propagation, spatial consistency, and symbolic deduction. Using a reinforcement lear"},"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":"2506.04821","kind":"arxiv","version":1},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.LG","submitted_at":"2025-06-05T09:40:47Z","cross_cats_sorted":[],"title_canon_sha256":"a4a13ea8cc97dd3bbd230299a21e01e9543c40a24c8d6dd2e48cf0604beba52f","abstract_canon_sha256":"09f3f85fc938e338d8c42789e3b92e032004073cfa983210ed7e91ca200330f5"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T11:16:37.254374Z","signature_b64":"nZbM6z1v46gfA0Lxel+MLnhJXgJUdred9DPZwi7Du7RPPwG9MxzwqG5CnMq+yyuFLpnOPOgpmbDIh7qPqer1BQ==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"07865a4b1b3acdb099b9f409c6236d56ed6c48695edfebc04f96dfcf19143174","last_reissued_at":"2026-07-05T11:16:37.253813Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T11:16:37.253813Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"LogicPuzzleRL: Cultivating Robust Mathematical Reasoning in LLMs via Reinforcement Learning","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":[],"primary_cat":"cs.LG","authors_text":"Bin Cui, Chengyu Shen, Hao Liang, Hejun Dong, Jingwen Deng, Qijie You, Runming He, Wentao Zhang, Zhen Hao Wong, Zirong Chen","submitted_at":"2025-06-05T09:40:47Z","abstract_excerpt":"Large language models (LLMs) excel at many supervised tasks but often struggle with structured reasoning in unfamiliar settings. This discrepancy suggests that standard fine-tuning pipelines may instill narrow, domain-specific heuristics rather than fostering general-purpose thinking strategies. In this work, we propose a \"play to learn\" framework that fine-tunes LLMs through reinforcement learning on a suite of seven custom logic puzzles, each designed to cultivate distinct reasoning skills such as constraint propagation, spatial consistency, and symbolic deduction. Using a reinforcement lear"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2506.04821","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/2506.04821/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":"2506.04821","created_at":"2026-07-05T11:16:37.253875+00:00"},{"alias_kind":"arxiv_version","alias_value":"2506.04821v1","created_at":"2026-07-05T11:16:37.253875+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2506.04821","created_at":"2026-07-05T11:16:37.253875+00:00"},{"alias_kind":"pith_short_12","alias_value":"A6DFUSY3HLG3","created_at":"2026-07-05T11:16:37.253875+00:00"},{"alias_kind":"pith_short_16","alias_value":"A6DFUSY3HLG3BGNZ","created_at":"2026-07-05T11:16:37.253875+00:00"},{"alias_kind":"pith_short_8","alias_value":"A6DFUSY3","created_at":"2026-07-05T11:16:37.253875+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":4,"internal_anchor_count":0,"sample":[{"citing_arxiv_id":"2606.05464","citing_title":"Step-by-Step Optimization-like Reasoning in LLMs over Expanding Search Spaces","ref_index":21,"is_internal_anchor":false},{"citing_arxiv_id":"2606.06556","citing_title":"Robots Need More than VLA and World Models","ref_index":150,"is_internal_anchor":false},{"citing_arxiv_id":"2507.04023","citing_title":"Do LLMs Overthink Basic Math Reasoning? Benchmarking the Accuracy-Efficiency Tradeoff in Language Models","ref_index":25,"is_internal_anchor":false},{"citing_arxiv_id":"2604.19773","citing_title":"PR-CAD: Progressive Refinement for Unified Controllable and Faithful Text-to-CAD Generation with Large Language Models","ref_index":17,"is_internal_anchor":false}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/A6DFUSY3HLG3BGNZ6QE4MI3NK3","json":"https://pith.science/pith/A6DFUSY3HLG3BGNZ6QE4MI3NK3.json","graph_json":"https://pith.science/api/pith-number/A6DFUSY3HLG3BGNZ6QE4MI3NK3/graph.json","events_json":"https://pith.science/api/pith-number/A6DFUSY3HLG3BGNZ6QE4MI3NK3/events.json","paper":"https://pith.science/paper/A6DFUSY3"},"agent_actions":{"view_html":"https://pith.science/pith/A6DFUSY3HLG3BGNZ6QE4MI3NK3","download_json":"https://pith.science/pith/A6DFUSY3HLG3BGNZ6QE4MI3NK3.json","view_paper":"https://pith.science/paper/A6DFUSY3","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2506.04821&json=true","fetch_graph":"https://pith.science/api/pith-number/A6DFUSY3HLG3BGNZ6QE4MI3NK3/graph.json","fetch_events":"https://pith.science/api/pith-number/A6DFUSY3HLG3BGNZ6QE4MI3NK3/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/A6DFUSY3HLG3BGNZ6QE4MI3NK3/action/timestamp_anchor","attest_storage":"https://pith.science/pith/A6DFUSY3HLG3BGNZ6QE4MI3NK3/action/storage_attestation","attest_author":"https://pith.science/pith/A6DFUSY3HLG3BGNZ6QE4MI3NK3/action/author_attestation","sign_citation":"https://pith.science/pith/A6DFUSY3HLG3BGNZ6QE4MI3NK3/action/citation_signature","submit_replication":"https://pith.science/pith/A6DFUSY3HLG3BGNZ6QE4MI3NK3/action/replication_record"}},"created_at":"2026-07-05T11:16:37.253875+00:00","updated_at":"2026-07-05T11:16:37.253875+00:00"}