{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2025:TKKPQ7VGF4FQ6KF3PIGUYON6Z5","short_pith_number":"pith:TKKPQ7VG","schema_version":"1.0","canonical_sha256":"9a94f87ea62f0b0f28bb7a0d4c39becf6512157f0fa73352978c436b1695acd6","source":{"kind":"arxiv","id":"2504.18587","version":1},"attestation_state":"computed","paper":{"title":"Training Large Language Models to Reason via EM Policy Gradient","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.AI","stat.ML"],"primary_cat":"cs.LG","authors_text":"Tianbing Xu","submitted_at":"2025-04-24T01:31:05Z","abstract_excerpt":"Recently, foundation models such as OpenAI's O1 and O3, along with DeepSeek's R1, have demonstrated strong reasoning capacities and problem-solving skills acquired through large-scale reinforcement learning (RL), with wide applications in mathematics, coding, science, intelligent agents, and virtual assistants. In this work, we introduce an off-policy reinforcement learning algorithm, EM Policy Gradient, aimed at enhancing LLM reasoning by optimizing expected return over reasoning trajectories. We frame the reasoning task as an Expectation-Maximization (EM) optimization problem, alternating be"},"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":"2504.18587","kind":"arxiv","version":1},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.LG","submitted_at":"2025-04-24T01:31:05Z","cross_cats_sorted":["cs.AI","stat.ML"],"title_canon_sha256":"a7852d851c0de32cb089099cbc0db4cd2851155bc42aa0ecb666c25d95f2b325","abstract_canon_sha256":"e9e37f6ebf66ac0ba1c462b7cca11bf0e1202c7332a55398885637bfdbe9e707"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T10:54:17.502334Z","signature_b64":"oSSUvOOuCRnrLDo3BpyCv3cjnK22aVkVK55K+oJwnT0pKOnESlckicoxYQkf+ZAzsSBsKe+fom/cYUlGwx7yBg==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"9a94f87ea62f0b0f28bb7a0d4c39becf6512157f0fa73352978c436b1695acd6","last_reissued_at":"2026-07-05T10:54:17.501845Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T10:54:17.501845Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Training Large Language Models to Reason via EM Policy Gradient","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.AI","stat.ML"],"primary_cat":"cs.LG","authors_text":"Tianbing Xu","submitted_at":"2025-04-24T01:31:05Z","abstract_excerpt":"Recently, foundation models such as OpenAI's O1 and O3, along with DeepSeek's R1, have demonstrated strong reasoning capacities and problem-solving skills acquired through large-scale reinforcement learning (RL), with wide applications in mathematics, coding, science, intelligent agents, and virtual assistants. In this work, we introduce an off-policy reinforcement learning algorithm, EM Policy Gradient, aimed at enhancing LLM reasoning by optimizing expected return over reasoning trajectories. We frame the reasoning task as an Expectation-Maximization (EM) optimization problem, alternating be"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2504.18587","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/2504.18587/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":"2504.18587","created_at":"2026-07-05T10:54:17.501905+00:00"},{"alias_kind":"arxiv_version","alias_value":"2504.18587v1","created_at":"2026-07-05T10:54:17.501905+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2504.18587","created_at":"2026-07-05T10:54:17.501905+00:00"},{"alias_kind":"pith_short_12","alias_value":"TKKPQ7VGF4FQ","created_at":"2026-07-05T10:54:17.501905+00:00"},{"alias_kind":"pith_short_16","alias_value":"TKKPQ7VGF4FQ6KF3","created_at":"2026-07-05T10:54:17.501905+00:00"},{"alias_kind":"pith_short_8","alias_value":"TKKPQ7VG","created_at":"2026-07-05T10:54:17.501905+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":1,"internal_anchor_count":1,"sample":[{"citing_arxiv_id":"2606.00726","citing_title":"Latent Reward Steering: An Adaptive Inference-Time Framework that Implicitly Promotes Cognitive Behaviors in Reasoning LLMs","ref_index":29,"is_internal_anchor":true}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/TKKPQ7VGF4FQ6KF3PIGUYON6Z5","json":"https://pith.science/pith/TKKPQ7VGF4FQ6KF3PIGUYON6Z5.json","graph_json":"https://pith.science/api/pith-number/TKKPQ7VGF4FQ6KF3PIGUYON6Z5/graph.json","events_json":"https://pith.science/api/pith-number/TKKPQ7VGF4FQ6KF3PIGUYON6Z5/events.json","paper":"https://pith.science/paper/TKKPQ7VG"},"agent_actions":{"view_html":"https://pith.science/pith/TKKPQ7VGF4FQ6KF3PIGUYON6Z5","download_json":"https://pith.science/pith/TKKPQ7VGF4FQ6KF3PIGUYON6Z5.json","view_paper":"https://pith.science/paper/TKKPQ7VG","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2504.18587&json=true","fetch_graph":"https://pith.science/api/pith-number/TKKPQ7VGF4FQ6KF3PIGUYON6Z5/graph.json","fetch_events":"https://pith.science/api/pith-number/TKKPQ7VGF4FQ6KF3PIGUYON6Z5/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/TKKPQ7VGF4FQ6KF3PIGUYON6Z5/action/timestamp_anchor","attest_storage":"https://pith.science/pith/TKKPQ7VGF4FQ6KF3PIGUYON6Z5/action/storage_attestation","attest_author":"https://pith.science/pith/TKKPQ7VGF4FQ6KF3PIGUYON6Z5/action/author_attestation","sign_citation":"https://pith.science/pith/TKKPQ7VGF4FQ6KF3PIGUYON6Z5/action/citation_signature","submit_replication":"https://pith.science/pith/TKKPQ7VGF4FQ6KF3PIGUYON6Z5/action/replication_record"}},"created_at":"2026-07-05T10:54:17.501905+00:00","updated_at":"2026-07-05T10:54:17.501905+00:00"}