{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2025:RSUHSOK5HKNMRZ36OH3SFALUZ3","short_pith_number":"pith:RSUHSOK5","schema_version":"1.0","canonical_sha256":"8ca879395d3a9ac8e77e71f7228174cefc54f5284bdc159e0193933016d2f171","source":{"kind":"arxiv","id":"2510.14807","version":3},"attestation_state":"computed","paper":{"title":"Beyond the Sampled Token: Preserving Candidate Support in RLVR","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"","cross_cats":[],"primary_cat":"cs.AI","authors_text":"Ruotian Peng, Weiyang Liu, Yandong Wen, Yi Ren, Zhouliang Yu","submitted_at":"2025-10-16T15:40:49Z","abstract_excerpt":"We revisit exploration collapse in reinforcement learning with verifiable rewards (RLVR), from the perspective of the \\emph{candidate distribution} for next-token prediction. We formally show that as probability concentrates on the top-$1$ candidate, the expected number of distinct responses collapses to one regardless of the sampling budget $K$. This theoretical implication is further verified by our empirical tracking of top-$N$ candidate probabilities during training, where the top-$1$ candidate progressively dominates while plausible alternatives are suppressed. These findings suggest a ke"},"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":"2510.14807","kind":"arxiv","version":3},"metadata":{"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.AI","submitted_at":"2025-10-16T15:40:49Z","cross_cats_sorted":[],"title_canon_sha256":"54019da39b72180d41e4009581faea96da4b51d98ed51ea992a20bfa08790e65","abstract_canon_sha256":"421185bca8cd9f0ec411e8b0dcfb550c93adbb5499966a26b2d7e45a56b97a45"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-06-19T16:09:51.661603Z","signature_b64":"hWIByFmWhshROApPNPAU92NF6aTDPG3XlSSnvt0huH8zGIip5T8IC3WBnPc3kYjT/zgAGVdy3Yr6/ajnQRClBg==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"8ca879395d3a9ac8e77e71f7228174cefc54f5284bdc159e0193933016d2f171","last_reissued_at":"2026-06-19T16:09:51.661164Z","signature_status":"signed_v1","first_computed_at":"2026-06-19T16:09:51.661164Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Beyond the Sampled Token: Preserving Candidate Support in RLVR","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"","cross_cats":[],"primary_cat":"cs.AI","authors_text":"Ruotian Peng, Weiyang Liu, Yandong Wen, Yi Ren, Zhouliang Yu","submitted_at":"2025-10-16T15:40:49Z","abstract_excerpt":"We revisit exploration collapse in reinforcement learning with verifiable rewards (RLVR), from the perspective of the \\emph{candidate distribution} for next-token prediction. We formally show that as probability concentrates on the top-$1$ candidate, the expected number of distinct responses collapses to one regardless of the sampling budget $K$. This theoretical implication is further verified by our empirical tracking of top-$N$ candidate probabilities during training, where the top-$1$ candidate progressively dominates while plausible alternatives are suppressed. These findings suggest a ke"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2510.14807","kind":"arxiv","version":3},"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/2510.14807/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":"2510.14807","created_at":"2026-06-19T16:09:51.661221+00:00"},{"alias_kind":"arxiv_version","alias_value":"2510.14807v3","created_at":"2026-06-19T16:09:51.661221+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2510.14807","created_at":"2026-06-19T16:09:51.661221+00:00"},{"alias_kind":"pith_short_12","alias_value":"RSUHSOK5HKNM","created_at":"2026-06-19T16:09:51.661221+00:00"},{"alias_kind":"pith_short_16","alias_value":"RSUHSOK5HKNMRZ36","created_at":"2026-06-19T16:09:51.661221+00:00"},{"alias_kind":"pith_short_8","alias_value":"RSUHSOK5","created_at":"2026-06-19T16:09:51.661221+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":9,"internal_anchor_count":9,"sample":[{"citing_arxiv_id":"2605.06094","citing_title":"VISD: Enhancing Video Reasoning via Structured Self-Distillation","ref_index":32,"is_internal_anchor":true},{"citing_arxiv_id":"2601.18832","citing_title":"The Geometric Reasoner: Manifold-Informed Latent Foresight Search for Long-Context Reasoning","ref_index":16,"is_internal_anchor":true},{"citing_arxiv_id":"2605.06094","citing_title":"VISD: Enhancing Video Reasoning via Structured Self-Distillation","ref_index":32,"is_internal_anchor":true},{"citing_arxiv_id":"2605.06094","citing_title":"VISD: Enhancing Video Reasoning via Structured Self-Distillation","ref_index":32,"is_internal_anchor":true},{"citing_arxiv_id":"2605.00380","citing_title":"ResRL: Boosting LLM Reasoning via Negative Sample Projection Residual Reinforcement Learning","ref_index":58,"is_internal_anchor":true},{"citing_arxiv_id":"2604.14142","citing_title":"From $P(y|x)$ to $P(y)$: Investigating Reinforcement Learning in Pre-train Space","ref_index":42,"is_internal_anchor":true},{"citing_arxiv_id":"2604.12632","citing_title":"Calibration-Aware Policy Optimization for Reasoning LLMs","ref_index":24,"is_internal_anchor":true},{"citing_arxiv_id":"2605.00380","citing_title":"ResRL: Boosting LLM Reasoning via Negative Sample Projection Residual Reinforcement Learning","ref_index":58,"is_internal_anchor":true},{"citing_arxiv_id":"2605.06094","citing_title":"VISD: Enhancing Video Reasoning via Structured Self-Distillation","ref_index":32,"is_internal_anchor":true}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/RSUHSOK5HKNMRZ36OH3SFALUZ3","json":"https://pith.science/pith/RSUHSOK5HKNMRZ36OH3SFALUZ3.json","graph_json":"https://pith.science/api/pith-number/RSUHSOK5HKNMRZ36OH3SFALUZ3/graph.json","events_json":"https://pith.science/api/pith-number/RSUHSOK5HKNMRZ36OH3SFALUZ3/events.json","paper":"https://pith.science/paper/RSUHSOK5"},"agent_actions":{"view_html":"https://pith.science/pith/RSUHSOK5HKNMRZ36OH3SFALUZ3","download_json":"https://pith.science/pith/RSUHSOK5HKNMRZ36OH3SFALUZ3.json","view_paper":"https://pith.science/paper/RSUHSOK5","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2510.14807&json=true","fetch_graph":"https://pith.science/api/pith-number/RSUHSOK5HKNMRZ36OH3SFALUZ3/graph.json","fetch_events":"https://pith.science/api/pith-number/RSUHSOK5HKNMRZ36OH3SFALUZ3/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/RSUHSOK5HKNMRZ36OH3SFALUZ3/action/timestamp_anchor","attest_storage":"https://pith.science/pith/RSUHSOK5HKNMRZ36OH3SFALUZ3/action/storage_attestation","attest_author":"https://pith.science/pith/RSUHSOK5HKNMRZ36OH3SFALUZ3/action/author_attestation","sign_citation":"https://pith.science/pith/RSUHSOK5HKNMRZ36OH3SFALUZ3/action/citation_signature","submit_replication":"https://pith.science/pith/RSUHSOK5HKNMRZ36OH3SFALUZ3/action/replication_record"}},"created_at":"2026-06-19T16:09:51.661221+00:00","updated_at":"2026-06-19T16:09:51.661221+00:00"}