{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2026:MXQRVVEXJCGCJPMFUNJNCJLYPA","short_pith_number":"pith:MXQRVVEX","schema_version":"1.0","canonical_sha256":"65e11ad497488c24bd85a352d12578782110e39f9949f52cac3178f8cbc9b70f","source":{"kind":"arxiv","id":"2601.15158","version":4},"attestation_state":"computed","paper":{"title":"Outcome-Based RL Provably Leads Transformers to Reason, but Only With the Right Data","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.AI"],"primary_cat":"cs.LG","authors_text":"Nadav Cohen, Shahar Mendel, Yotam Alexander, Yuval Ran-Milo","submitted_at":"2026-01-21T16:36:19Z","abstract_excerpt":"Transformers trained via Reinforcement Learning (RL) with outcome-based supervision can spontaneously develop the ability to generate intermediate reasoning steps (Chain-of-Thought). Yet the mechanism by which sparse rewards drive policy gradient to discover such systematic reasoning remains poorly understood. We address this by analyzing the policy gradient dynamics of single-layer Transformers on a synthetic graph traversal task that cannot be solved without Chain-of-Thought but admits a simple iterative solution. We prove that despite training solely on final-answer correctness, policy grad"},"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":"2601.15158","kind":"arxiv","version":4},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.LG","submitted_at":"2026-01-21T16:36:19Z","cross_cats_sorted":["cs.AI"],"title_canon_sha256":"bf80f8176593950e78c33aa3a2347d8ef5c7690b2869462eeba80a1522c5064a","abstract_canon_sha256":"19bebb70abcc094017accc4ca61f1e4e87e7646cabe32d4880b5270d3512afd2"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-06-04T01:08:40.290097Z","signature_b64":"7R06nSD0ATeesl6/lmmHEdhWNq1FAo3BlgT2eTR85jT2c2c+QhMeKAcP1p1CMWr3vSkXxQazmbmDNxaz364MDA==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"65e11ad497488c24bd85a352d12578782110e39f9949f52cac3178f8cbc9b70f","last_reissued_at":"2026-06-04T01:08:40.289270Z","signature_status":"signed_v1","first_computed_at":"2026-06-04T01:08:40.289270Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Outcome-Based RL Provably Leads Transformers to Reason, but Only With the Right Data","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.AI"],"primary_cat":"cs.LG","authors_text":"Nadav Cohen, Shahar Mendel, Yotam Alexander, Yuval Ran-Milo","submitted_at":"2026-01-21T16:36:19Z","abstract_excerpt":"Transformers trained via Reinforcement Learning (RL) with outcome-based supervision can spontaneously develop the ability to generate intermediate reasoning steps (Chain-of-Thought). Yet the mechanism by which sparse rewards drive policy gradient to discover such systematic reasoning remains poorly understood. We address this by analyzing the policy gradient dynamics of single-layer Transformers on a synthetic graph traversal task that cannot be solved without Chain-of-Thought but admits a simple iterative solution. We prove that despite training solely on final-answer correctness, policy grad"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2601.15158","kind":"arxiv","version":4},"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/2601.15158/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":"2601.15158","created_at":"2026-06-04T01:08:40.289389+00:00"},{"alias_kind":"arxiv_version","alias_value":"2601.15158v4","created_at":"2026-06-04T01:08:40.289389+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2601.15158","created_at":"2026-06-04T01:08:40.289389+00:00"},{"alias_kind":"pith_short_12","alias_value":"MXQRVVEXJCGC","created_at":"2026-06-04T01:08:40.289389+00:00"},{"alias_kind":"pith_short_16","alias_value":"MXQRVVEXJCGCJPMF","created_at":"2026-06-04T01:08:40.289389+00:00"},{"alias_kind":"pith_short_8","alias_value":"MXQRVVEX","created_at":"2026-06-04T01:08:40.289389+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":2,"internal_anchor_count":2,"sample":[{"citing_arxiv_id":"2604.25872","citing_title":"When Errors Can Be Beneficial: A Categorization of Imperfect Rewards for Policy Gradient","ref_index":63,"is_internal_anchor":true},{"citing_arxiv_id":"2605.06992","citing_title":"Why Does Agentic Safety Fail to Generalize Across Tasks?","ref_index":92,"is_internal_anchor":true}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/MXQRVVEXJCGCJPMFUNJNCJLYPA","json":"https://pith.science/pith/MXQRVVEXJCGCJPMFUNJNCJLYPA.json","graph_json":"https://pith.science/api/pith-number/MXQRVVEXJCGCJPMFUNJNCJLYPA/graph.json","events_json":"https://pith.science/api/pith-number/MXQRVVEXJCGCJPMFUNJNCJLYPA/events.json","paper":"https://pith.science/paper/MXQRVVEX"},"agent_actions":{"view_html":"https://pith.science/pith/MXQRVVEXJCGCJPMFUNJNCJLYPA","download_json":"https://pith.science/pith/MXQRVVEXJCGCJPMFUNJNCJLYPA.json","view_paper":"https://pith.science/paper/MXQRVVEX","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2601.15158&json=true","fetch_graph":"https://pith.science/api/pith-number/MXQRVVEXJCGCJPMFUNJNCJLYPA/graph.json","fetch_events":"https://pith.science/api/pith-number/MXQRVVEXJCGCJPMFUNJNCJLYPA/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/MXQRVVEXJCGCJPMFUNJNCJLYPA/action/timestamp_anchor","attest_storage":"https://pith.science/pith/MXQRVVEXJCGCJPMFUNJNCJLYPA/action/storage_attestation","attest_author":"https://pith.science/pith/MXQRVVEXJCGCJPMFUNJNCJLYPA/action/author_attestation","sign_citation":"https://pith.science/pith/MXQRVVEXJCGCJPMFUNJNCJLYPA/action/citation_signature","submit_replication":"https://pith.science/pith/MXQRVVEXJCGCJPMFUNJNCJLYPA/action/replication_record"}},"created_at":"2026-06-04T01:08:40.289389+00:00","updated_at":"2026-06-04T01:08:40.289389+00:00"}