{"state_type":"pith_open_graph_state","state_version":"1.0","pith_number":"pith:2025:Q7OXT2D7TV2U5BL2A5ITHAAZGE","merge_version":"pith-open-graph-merge-v1","event_count":2,"valid_event_count":2,"invalid_event_count":0,"equivocation_count":0,"current":{"canonical_record":{"metadata":{"abstract_canon_sha256":"99e9eb6e20d4e54ae62f41a827a7d314d315ad5ae79695f60a26a1bedff501b7","cross_cats_sorted":["cs.AI"],"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.CL","submitted_at":"2025-02-20T17:49:26Z","title_canon_sha256":"58ca94316949335c21db4792a681216bb1a96c1f2781187232e95607a0904f69"},"schema_version":"1.0","source":{"id":"2502.14768","kind":"arxiv","version":1}},"source_aliases":[{"alias_kind":"arxiv","alias_value":"2502.14768","created_at":"2026-05-17T23:38:46Z"},{"alias_kind":"arxiv_version","alias_value":"2502.14768v1","created_at":"2026-05-17T23:38:46Z"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2502.14768","created_at":"2026-05-17T23:38:46Z"},{"alias_kind":"pith_short_12","alias_value":"Q7OXT2D7TV2U","created_at":"2026-05-18T12:33:37Z"},{"alias_kind":"pith_short_16","alias_value":"Q7OXT2D7TV2U5BL2","created_at":"2026-05-18T12:33:37Z"},{"alias_kind":"pith_short_8","alias_value":"Q7OXT2D7","created_at":"2026-05-18T12:33:37Z"}],"graph_snapshots":[{"event_id":"sha256:fc9ddd469c7dced98bc45a4f5c53c7d7bea35a53ada0308807276c4e6ab99933","target":"graph","created_at":"2026-05-17T23:38:46Z","signer":{"key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54","signer_id":"pith.science","signer_type":"pith_registry"},"payload":{"graph_snapshot":{"author_claims":{"count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57","strong_count":0},"builder_version":"pith-number-builder-2026-05-17-v1","claims":{"count":4,"items":[{"attestation":"unclaimed","claim_id":"C1","kind":"strongest_claim","source":"verdict.strongest_claim","status":"machine_extracted","text":"after training on just 5K logic problems, it demonstrates generalization abilities to the challenging math benchmarks AIME and AMC."},{"attestation":"unclaimed","claim_id":"C2","kind":"weakest_assumption","source":"verdict.weakest_assumption","status":"machine_extracted","text":"That the advanced reasoning behaviors (reflection, verification, summarization) are induced by the RL process rather than already latent in the base 7B model or triggered by the system prompt alone."},{"attestation":"unclaimed","claim_id":"C3","kind":"one_line_summary","source":"verdict.one_line_summary","status":"machine_extracted","text":"Rule-based RL on 5K logic puzzles induces advanced reasoning in a 7B model that transfers to AIME and AMC."},{"attestation":"unclaimed","claim_id":"C4","kind":"headline","source":"verdict.pith_extraction.headline","status":"machine_extracted","text":"Rule-based RL on 5K logic puzzles induces reflection and verification in a 7B model that transfers to AIME and AMC."}],"snapshot_sha256":"0d6704e1857d5c09f9ba6ab79da66c16486e1a231d2bfd5d7fdedfa928e0cbc5"},"formal_canon":{"evidence_count":2,"snapshot_sha256":"0e90cfa9e1b05ca97a7f7e4f53b66aa44d77108040460b51d770bceb21ae6745"},"paper":{"abstract_excerpt":"Inspired by the success of DeepSeek-R1, we explore the potential of rule-based reinforcement learning (RL) in large reasoning models. To analyze reasoning dynamics, we use synthetic logic puzzles as training data due to their controllable complexity and straightforward answer verification. We make some key technical contributions that lead to effective and stable RL training: a system prompt that emphasizes the thinking and answering process, a stringent format reward function that penalizes outputs for taking shortcuts, and a straightforward training recipe that achieves stable convergence. O","authors_text":"Bryan Dai, Chong Luo, Haoming Luo, Joey Zhou, Kai Qiu, Qingnan Ren, Tian Xie, Yuqian Hong, Zhirong Wu, Zitian Gao","cross_cats":["cs.AI"],"headline":"Rule-based RL on 5K logic puzzles induces reflection and verification in a 7B model that transfers to AIME and AMC.","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.CL","submitted_at":"2025-02-20T17:49:26Z","title":"Logic-RL: Unleashing LLM Reasoning with Rule-Based Reinforcement Learning"},"references":{"count":27,"internal_anchors":0,"resolved_work":27,"sample":[{"cited_arxiv_id":"","doi":"","is_internal_anchor":false,"ref_index":1,"title":"Le, Sergey Levine, and Yi Ma","work_id":"ad2bcec1-178f-4281-a602-c683a5711e90","year":2025},{"cited_arxiv_id":"","doi":"","is_internal_anchor":false,"ref_index":2,"title":"Training verifiers to solve math word problems","work_id":"1f7ff91e-3d16-4cba-ad18-3c6ae7ec674b","year":2021},{"cited_arxiv_id":"","doi":"","is_internal_anchor":false,"ref_index":3,"title":"DeepSeek-AI, Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Ruoyu Zhang, Runxin Xu, Qihao Zhu, Shirong Ma, Peiyi Wang, Xiao Bi, Xiaokang Zhang, Xingkai Yu, Yu Wu, Z. F. Wu, Zhibin Gou, Zhihong Sha","work_id":"2472dd86-1219-472e-af5b-4fba94816f78","year":2025},{"cited_arxiv_id":"","doi":"","is_internal_anchor":false,"ref_index":4,"title":"Alphazero-like tree-search can guide large language model decoding and training, 2024","work_id":"b23b3d7c-5200-48e4-9a0c-344903052311","year":2024},{"cited_arxiv_id":"","doi":"","is_internal_anchor":false,"ref_index":5,"title":"Omni-math: A universal olympiad level mathematic benchmark for large language models, 2024","work_id":"53e56ad1-de9e-460f-8183-689e5f4b23d6","year":2024}],"snapshot_sha256":"bf59d0bbcfb1fdd16bbf679f7ba25ea9e77e3bff305477f0e0b128e2384651b7"},"source":{"id":"2502.14768","kind":"arxiv","version":1},"verdict":{"created_at":"2026-05-16T20:54:31.737562Z","id":"1f47f35f-6ebc-4a36-8323-0b024479609f","model_set":{"reader":"grok-4.3"},"one_line_summary":"Rule-based RL on 5K logic puzzles induces advanced reasoning in a 7B model that transfers to AIME and AMC.","pipeline_version":"pith-pipeline@v0.9.0","pith_extraction_headline":"Rule-based RL on 5K logic puzzles induces reflection and verification in a 7B model that transfers to AIME and AMC.","strongest_claim":"after training on just 5K logic problems, it demonstrates generalization abilities to the challenging math benchmarks AIME and AMC.","weakest_assumption":"That the advanced reasoning behaviors (reflection, verification, summarization) are induced by the RL process rather than already latent in the base 7B model or triggered by the system prompt alone."}},"verdict_id":"1f47f35f-6ebc-4a36-8323-0b024479609f"}}],"author_attestations":[],"timestamp_anchors":[],"storage_attestations":[],"citation_signatures":[],"replication_records":[],"corrections":[],"mirror_hints":[],"record_created":{"event_id":"sha256:b7d0189b737de2cc05425ca54736d9b9714054bdd4f7ce73c3816cd9f656ae5a","target":"record","created_at":"2026-05-17T23:38:46Z","signer":{"key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54","signer_id":"pith.science","signer_type":"pith_registry"},"payload":{"attestation_state":"computed","canonical_record":{"metadata":{"abstract_canon_sha256":"99e9eb6e20d4e54ae62f41a827a7d314d315ad5ae79695f60a26a1bedff501b7","cross_cats_sorted":["cs.AI"],"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.CL","submitted_at":"2025-02-20T17:49:26Z","title_canon_sha256":"58ca94316949335c21db4792a681216bb1a96c1f2781187232e95607a0904f69"},"schema_version":"1.0","source":{"id":"2502.14768","kind":"arxiv","version":1}},"canonical_sha256":"87dd79e87f9d754e857a0751338019311f6aa80cf62ca2dad3f15188522b86b2","receipt":{"algorithm":"ed25519","builder_version":"pith-number-builder-2026-05-17-v1","canonical_sha256":"87dd79e87f9d754e857a0751338019311f6aa80cf62ca2dad3f15188522b86b2","first_computed_at":"2026-05-17T23:38:46.595890Z","key_id":"pith-v1-2026-05","kind":"pith_receipt","last_reissued_at":"2026-05-17T23:38:46.595890Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54","receipt_version":"0.3","signature_b64":"L267w2kGeCg1d97CPeeBq0m0kBwVJCuoCjEt7BiW87+1QsPUi2Dl2vDD0EkrRzzmFm2ucBtBODOKXgvkBL2MDQ==","signature_status":"signed_v1","signed_at":"2026-05-17T23:38:46.596356Z","signed_message":"canonical_sha256_bytes"},"source_id":"2502.14768","source_kind":"arxiv","source_version":1}}},"equivocations":[],"invalid_events":[],"applied_event_ids":["sha256:b7d0189b737de2cc05425ca54736d9b9714054bdd4f7ce73c3816cd9f656ae5a","sha256:fc9ddd469c7dced98bc45a4f5c53c7d7bea35a53ada0308807276c4e6ab99933"],"state_sha256":"4afffec4d7ad435a70ee732bde0fc9c5f42407d1026afc57d750bdd647970793"}