{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2025:SL4E4M2VKVFGAOXSBOVYPKP65W","short_pith_number":"pith:SL4E4M2V","schema_version":"1.0","canonical_sha256":"92f84e3355554a603af20bab87a9feeda13c546013136ed2654de88e19bc9555","source":{"kind":"arxiv","id":"2506.07468","version":4},"attestation_state":"computed","paper":{"title":"Chasing Moving Targets with Online Self-Play Reinforcement Learning for Safer Language Models","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.CL","cs.MA"],"primary_cat":"cs.LG","authors_text":"Liwei Jiang, Mickel Liu, Natasha Jaques, Simon Shaolei Du, Tim Althoff, Yancheng Liang, Yejin Choi","submitted_at":"2025-06-09T06:35:12Z","abstract_excerpt":"Conventional large language model (LLM) safety alignment relies on a reactive, disjoint loop: attackers exploit a static model, then defenders patch exposed vulnerabilities. This sequential setup leads to attackers overfitting obsolete exploits while defenders perpetually lag behind emerging threats. To address this, we introduce Self-RedTeam, the first fully online self-play multi-agent reinforcement learning (MARL) algorithm that continuously co-evolves attacker and defender for robust safety alignment. A single policy self-plays as both attacker and defender, generating adversarial prompts "},"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":"2506.07468","kind":"arxiv","version":4},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.LG","submitted_at":"2025-06-09T06:35:12Z","cross_cats_sorted":["cs.CL","cs.MA"],"title_canon_sha256":"59dcaa83c310d94b5c705067f8b34392f164769b2426144d1d6b2ef5127ccc9f","abstract_canon_sha256":"7468fafae6b5bf0644764037ea1665103baf78694ff0ff46ae95f5fdb7a7963f"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-07T02:19:38.717468Z","signature_b64":"shVT4s4t598gsM33/zISNgvviPTBU1Us9YApeomghN2XlUYpXTu1ZEzp0+AAMy3GWi0wZJqcRJfbnDYL+A2sAQ==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"92f84e3355554a603af20bab87a9feeda13c546013136ed2654de88e19bc9555","last_reissued_at":"2026-07-07T02:19:38.716145Z","signature_status":"signed_v1","first_computed_at":"2026-07-07T02:19:38.716145Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Chasing Moving Targets with Online Self-Play Reinforcement Learning for Safer Language Models","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.CL","cs.MA"],"primary_cat":"cs.LG","authors_text":"Liwei Jiang, Mickel Liu, Natasha Jaques, Simon Shaolei Du, Tim Althoff, Yancheng Liang, Yejin Choi","submitted_at":"2025-06-09T06:35:12Z","abstract_excerpt":"Conventional large language model (LLM) safety alignment relies on a reactive, disjoint loop: attackers exploit a static model, then defenders patch exposed vulnerabilities. This sequential setup leads to attackers overfitting obsolete exploits while defenders perpetually lag behind emerging threats. To address this, we introduce Self-RedTeam, the first fully online self-play multi-agent reinforcement learning (MARL) algorithm that continuously co-evolves attacker and defender for robust safety alignment. A single policy self-plays as both attacker and defender, generating adversarial prompts "},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2506.07468","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/2506.07468/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":"2506.07468","created_at":"2026-07-07T02:19:38.716309+00:00"},{"alias_kind":"arxiv_version","alias_value":"2506.07468v4","created_at":"2026-07-07T02:19:38.716309+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2506.07468","created_at":"2026-07-07T02:19:38.716309+00:00"},{"alias_kind":"pith_short_12","alias_value":"SL4E4M2VKVFG","created_at":"2026-07-07T02:19:38.716309+00:00"},{"alias_kind":"pith_short_16","alias_value":"SL4E4M2VKVFGAOXS","created_at":"2026-07-07T02:19:38.716309+00:00"},{"alias_kind":"pith_short_8","alias_value":"SL4E4M2V","created_at":"2026-07-07T02:19:38.716309+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":6,"internal_anchor_count":6,"sample":[{"citing_arxiv_id":"2606.17657","citing_title":"Using Cognitive Models to Improve Language Model Simulation of Human Persuasion Games","ref_index":38,"is_internal_anchor":true},{"citing_arxiv_id":"2606.31748","citing_title":"Addressing Over-Refusal in LLMs with Competing Rewards","ref_index":103,"is_internal_anchor":true},{"citing_arxiv_id":"2504.09006","citing_title":"Learning in Structured Stackelberg Games","ref_index":34,"is_internal_anchor":true},{"citing_arxiv_id":"2605.08427","citing_title":"The Attacker in the Mirror: Breaking Self-Consistency in Safety via Anchored Bipolicy Self-Play","ref_index":22,"is_internal_anchor":true},{"citing_arxiv_id":"2604.24020","citing_title":"Poster: ClawdGo: Endogenous Security Awareness Training for Autonomous AI Agents","ref_index":11,"is_internal_anchor":true},{"citing_arxiv_id":"2605.01899","citing_title":"Disentangling Intent from Role: Adversarial Self-Play for Persona-Invariant Safety Alignment","ref_index":33,"is_internal_anchor":true}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/SL4E4M2VKVFGAOXSBOVYPKP65W","json":"https://pith.science/pith/SL4E4M2VKVFGAOXSBOVYPKP65W.json","graph_json":"https://pith.science/api/pith-number/SL4E4M2VKVFGAOXSBOVYPKP65W/graph.json","events_json":"https://pith.science/api/pith-number/SL4E4M2VKVFGAOXSBOVYPKP65W/events.json","paper":"https://pith.science/paper/SL4E4M2V"},"agent_actions":{"view_html":"https://pith.science/pith/SL4E4M2VKVFGAOXSBOVYPKP65W","download_json":"https://pith.science/pith/SL4E4M2VKVFGAOXSBOVYPKP65W.json","view_paper":"https://pith.science/paper/SL4E4M2V","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2506.07468&json=true","fetch_graph":"https://pith.science/api/pith-number/SL4E4M2VKVFGAOXSBOVYPKP65W/graph.json","fetch_events":"https://pith.science/api/pith-number/SL4E4M2VKVFGAOXSBOVYPKP65W/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/SL4E4M2VKVFGAOXSBOVYPKP65W/action/timestamp_anchor","attest_storage":"https://pith.science/pith/SL4E4M2VKVFGAOXSBOVYPKP65W/action/storage_attestation","attest_author":"https://pith.science/pith/SL4E4M2VKVFGAOXSBOVYPKP65W/action/author_attestation","sign_citation":"https://pith.science/pith/SL4E4M2VKVFGAOXSBOVYPKP65W/action/citation_signature","submit_replication":"https://pith.science/pith/SL4E4M2VKVFGAOXSBOVYPKP65W/action/replication_record"}},"created_at":"2026-07-07T02:19:38.716309+00:00","updated_at":"2026-07-07T02:19:38.716309+00:00"}