{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2025:QIJHMMIU7WIBIPJI5KXMXHCIXX","short_pith_number":"pith:QIJHMMIU","schema_version":"1.0","canonical_sha256":"8212763114fd90143d28eaaecb9c48bdf603ed0f57db8c36f39a4fa78f754446","source":{"kind":"arxiv","id":"2506.10406","version":1},"attestation_state":"computed","paper":{"title":"PAG: Multi-Turn Reinforced LLM Self-Correction with Policy as Generative Verifier","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.AI","cs.LG"],"primary_cat":"cs.CL","authors_text":"Chao Xin, Lin Yan, Qianchuan Zhao, Wenyuan Xu, Yufeng Yuan, Yuhua Jiang, Yuwen Xiong, Yu Yue","submitted_at":"2025-06-12T06:59:35Z","abstract_excerpt":"Large Language Models (LLMs) have demonstrated impressive capabilities in complex reasoning tasks, yet they still struggle to reliably verify the correctness of their own outputs. Existing solutions to this verification challenge often depend on separate verifier models or require multi-stage self-correction training pipelines, which limit scalability. In this paper, we propose Policy as Generative Verifier (PAG), a simple and effective framework that empowers LLMs to self-correct by alternating between policy and verifier roles within a unified multi-turn reinforcement learning (RL) paradigm."},"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":"2506.10406","kind":"arxiv","version":1},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.CL","submitted_at":"2025-06-12T06:59:35Z","cross_cats_sorted":["cs.AI","cs.LG"],"title_canon_sha256":"f2e9bcc801555309351daa46d2c058bc54005f1e1e3b7b03e39dd1fe91749451","abstract_canon_sha256":"2fb0f35ee56e3e2bf9349385c9dbd1bb5b6264ea0efc91c77499dd72860715dd"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T11:20:22.901537Z","signature_b64":"Gt4nALKMdvvDwcoigCRUqXg4uBNARJMlrIAhNuswGnwt35Vd6CYLT1c5+tqQGM6I7xgcbwBOEyY1FWGE8GSMDw==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"8212763114fd90143d28eaaecb9c48bdf603ed0f57db8c36f39a4fa78f754446","last_reissued_at":"2026-07-05T11:20:22.901018Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T11:20:22.901018Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"PAG: Multi-Turn Reinforced LLM Self-Correction with Policy as Generative Verifier","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.AI","cs.LG"],"primary_cat":"cs.CL","authors_text":"Chao Xin, Lin Yan, Qianchuan Zhao, Wenyuan Xu, Yufeng Yuan, Yuhua Jiang, Yuwen Xiong, Yu Yue","submitted_at":"2025-06-12T06:59:35Z","abstract_excerpt":"Large Language Models (LLMs) have demonstrated impressive capabilities in complex reasoning tasks, yet they still struggle to reliably verify the correctness of their own outputs. Existing solutions to this verification challenge often depend on separate verifier models or require multi-stage self-correction training pipelines, which limit scalability. In this paper, we propose Policy as Generative Verifier (PAG), a simple and effective framework that empowers LLMs to self-correct by alternating between policy and verifier roles within a unified multi-turn reinforcement learning (RL) paradigm."},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2506.10406","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/2506.10406/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":"2506.10406","created_at":"2026-07-05T11:20:22.901079+00:00"},{"alias_kind":"arxiv_version","alias_value":"2506.10406v1","created_at":"2026-07-05T11:20:22.901079+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2506.10406","created_at":"2026-07-05T11:20:22.901079+00:00"},{"alias_kind":"pith_short_12","alias_value":"QIJHMMIU7WIB","created_at":"2026-07-05T11:20:22.901079+00:00"},{"alias_kind":"pith_short_16","alias_value":"QIJHMMIU7WIBIPJI","created_at":"2026-07-05T11:20:22.901079+00:00"},{"alias_kind":"pith_short_8","alias_value":"QIJHMMIU","created_at":"2026-07-05T11:20:22.901079+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":6,"internal_anchor_count":0,"sample":[{"citing_arxiv_id":"2606.18910","citing_title":"REVES: REvision and VErification--Augmented Training for Test-Time Scaling","ref_index":51,"is_internal_anchor":false},{"citing_arxiv_id":"2606.01249","citing_title":"Trust Region On-Policy Distillation","ref_index":168,"is_internal_anchor":false},{"citing_arxiv_id":"2605.18851","citing_title":"STRIDE: Learnable Stepwise Language Feedback for LLM Reasoning","ref_index":30,"is_internal_anchor":false},{"citing_arxiv_id":"2509.18847","citing_title":"Failure Makes the Agent Stronger: Enhancing Accuracy through Structured Reflection for Reliable Tool Interactions","ref_index":28,"is_internal_anchor":false},{"citing_arxiv_id":"2509.08827","citing_title":"A Survey of Reinforcement Learning for Large Reasoning Models","ref_index":241,"is_internal_anchor":false},{"citing_arxiv_id":"2604.19656","citing_title":"Pause or Fabricate? Training Language Models for Grounded Reasoning","ref_index":14,"is_internal_anchor":false}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/QIJHMMIU7WIBIPJI5KXMXHCIXX","json":"https://pith.science/pith/QIJHMMIU7WIBIPJI5KXMXHCIXX.json","graph_json":"https://pith.science/api/pith-number/QIJHMMIU7WIBIPJI5KXMXHCIXX/graph.json","events_json":"https://pith.science/api/pith-number/QIJHMMIU7WIBIPJI5KXMXHCIXX/events.json","paper":"https://pith.science/paper/QIJHMMIU"},"agent_actions":{"view_html":"https://pith.science/pith/QIJHMMIU7WIBIPJI5KXMXHCIXX","download_json":"https://pith.science/pith/QIJHMMIU7WIBIPJI5KXMXHCIXX.json","view_paper":"https://pith.science/paper/QIJHMMIU","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2506.10406&json=true","fetch_graph":"https://pith.science/api/pith-number/QIJHMMIU7WIBIPJI5KXMXHCIXX/graph.json","fetch_events":"https://pith.science/api/pith-number/QIJHMMIU7WIBIPJI5KXMXHCIXX/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/QIJHMMIU7WIBIPJI5KXMXHCIXX/action/timestamp_anchor","attest_storage":"https://pith.science/pith/QIJHMMIU7WIBIPJI5KXMXHCIXX/action/storage_attestation","attest_author":"https://pith.science/pith/QIJHMMIU7WIBIPJI5KXMXHCIXX/action/author_attestation","sign_citation":"https://pith.science/pith/QIJHMMIU7WIBIPJI5KXMXHCIXX/action/citation_signature","submit_replication":"https://pith.science/pith/QIJHMMIU7WIBIPJI5KXMXHCIXX/action/replication_record"}},"created_at":"2026-07-05T11:20:22.901079+00:00","updated_at":"2026-07-05T11:20:22.901079+00:00"}