{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2025:LYA2NSAFTESK7RUQLR2WERHYEZ","short_pith_number":"pith:LYA2NSAF","schema_version":"1.0","canonical_sha256":"5e01a6c8059924afc6905c756244f8265bfa57fb3422afdff9c669a712111865","source":{"kind":"arxiv","id":"2503.19618","version":2},"attestation_state":"computed","paper":{"title":"Beyond Verifiable Rewards: Scaling Reinforcement Learning for Language Models to Unverifiable Data","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"","cross_cats":[],"primary_cat":"cs.LG","authors_text":"Lovish Madaan, R\\'emi Munos, Sid Wang, Yunhao Tang","submitted_at":"2025-03-25T13:03:09Z","abstract_excerpt":"We propose to scale RL to unverifiable data with a novel algorithm JEPO (Jensen's Evidence lower bound Policy Optimization). While most prior efforts on scaling RL for LLMs focus on verifiable data where ground truth answers are typically short-form and can be matched easily; we investigate the case where such assumptions are less valid (e.g., when answers are long-form such as mathematical proofs). To scale RL training to unverifiable data with contemporary training constraints, we propose JEPO. JEPO applies Jensen's evidence lower bound, a pragmatic simplification of the evidence lower bound"},"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":"2503.19618","kind":"arxiv","version":2},"metadata":{"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.LG","submitted_at":"2025-03-25T13:03:09Z","cross_cats_sorted":[],"title_canon_sha256":"4be6f49c9d013e4b4bc84e95f17e73b37d2418715feb236f33c05958f9c6f933","abstract_canon_sha256":"e63abfc126789fbe45cdc0b9e889da8c76f9adfe07e8ec7b4c3bf593e3a32064"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T11:11:21.904541Z","signature_b64":"ZOdME6fsJeOtqOpgDdFIlWANR281Z4d1T/6UG3fzAjBKGSvZHQSTmXP8cNaGEmBAbV41Pw4VFGfYhPRcnKjpAg==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"5e01a6c8059924afc6905c756244f8265bfa57fb3422afdff9c669a712111865","last_reissued_at":"2026-07-05T11:11:21.903985Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T11:11:21.903985Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Beyond Verifiable Rewards: Scaling Reinforcement Learning for Language Models to Unverifiable Data","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"","cross_cats":[],"primary_cat":"cs.LG","authors_text":"Lovish Madaan, R\\'emi Munos, Sid Wang, Yunhao Tang","submitted_at":"2025-03-25T13:03:09Z","abstract_excerpt":"We propose to scale RL to unverifiable data with a novel algorithm JEPO (Jensen's Evidence lower bound Policy Optimization). While most prior efforts on scaling RL for LLMs focus on verifiable data where ground truth answers are typically short-form and can be matched easily; we investigate the case where such assumptions are less valid (e.g., when answers are long-form such as mathematical proofs). To scale RL training to unverifiable data with contemporary training constraints, we propose JEPO. JEPO applies Jensen's evidence lower bound, a pragmatic simplification of the evidence lower bound"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2503.19618","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/2503.19618/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":"2503.19618","created_at":"2026-07-05T11:11:21.904044+00:00"},{"alias_kind":"arxiv_version","alias_value":"2503.19618v2","created_at":"2026-07-05T11:11:21.904044+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2503.19618","created_at":"2026-07-05T11:11:21.904044+00:00"},{"alias_kind":"pith_short_12","alias_value":"LYA2NSAFTESK","created_at":"2026-07-05T11:11:21.904044+00:00"},{"alias_kind":"pith_short_16","alias_value":"LYA2NSAFTESK7RUQ","created_at":"2026-07-05T11:11:21.904044+00:00"},{"alias_kind":"pith_short_8","alias_value":"LYA2NSAF","created_at":"2026-07-05T11:11:21.904044+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":5,"internal_anchor_count":0,"sample":[{"citing_arxiv_id":"2606.18089","citing_title":"From Reasoning Traces to Reusable Modules: Understanding Compositional Generalization in Language Model Reasoning","ref_index":63,"is_internal_anchor":false},{"citing_arxiv_id":"2605.10810","citing_title":"Likelihood scoring for continuations of mathematical text: a self-supervised benchmark with tests for shortcut vulnerabilities","ref_index":36,"is_internal_anchor":false},{"citing_arxiv_id":"2506.13351","citing_title":"Direct Reasoning Optimization: Token-Level Reasoning Reflectivity Meets Rubric Gates for Unverifiable Tasks","ref_index":27,"is_internal_anchor":false},{"citing_arxiv_id":"2509.14234","citing_title":"Compute as Teacher: Turning Inference Compute Into Reference-Free Supervision","ref_index":23,"is_internal_anchor":false},{"citing_arxiv_id":"2604.10110","citing_title":"Trust Your Memory: Verifiable Control of Smart Homes through Reinforcement Learning with Multi-dimensional Rewards","ref_index":17,"is_internal_anchor":false}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/LYA2NSAFTESK7RUQLR2WERHYEZ","json":"https://pith.science/pith/LYA2NSAFTESK7RUQLR2WERHYEZ.json","graph_json":"https://pith.science/api/pith-number/LYA2NSAFTESK7RUQLR2WERHYEZ/graph.json","events_json":"https://pith.science/api/pith-number/LYA2NSAFTESK7RUQLR2WERHYEZ/events.json","paper":"https://pith.science/paper/LYA2NSAF"},"agent_actions":{"view_html":"https://pith.science/pith/LYA2NSAFTESK7RUQLR2WERHYEZ","download_json":"https://pith.science/pith/LYA2NSAFTESK7RUQLR2WERHYEZ.json","view_paper":"https://pith.science/paper/LYA2NSAF","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2503.19618&json=true","fetch_graph":"https://pith.science/api/pith-number/LYA2NSAFTESK7RUQLR2WERHYEZ/graph.json","fetch_events":"https://pith.science/api/pith-number/LYA2NSAFTESK7RUQLR2WERHYEZ/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/LYA2NSAFTESK7RUQLR2WERHYEZ/action/timestamp_anchor","attest_storage":"https://pith.science/pith/LYA2NSAFTESK7RUQLR2WERHYEZ/action/storage_attestation","attest_author":"https://pith.science/pith/LYA2NSAFTESK7RUQLR2WERHYEZ/action/author_attestation","sign_citation":"https://pith.science/pith/LYA2NSAFTESK7RUQLR2WERHYEZ/action/citation_signature","submit_replication":"https://pith.science/pith/LYA2NSAFTESK7RUQLR2WERHYEZ/action/replication_record"}},"created_at":"2026-07-05T11:11:21.904044+00:00","updated_at":"2026-07-05T11:11:21.904044+00:00"}