{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2025:HT4L5CF2RXGKZJ5V275S3B26ID","short_pith_number":"pith:HT4L5CF2","schema_version":"1.0","canonical_sha256":"3cf8be88ba8dccaca7b5d7fb2d875e40c14dde51753bdf4e2f008febe1ca32f0","source":{"kind":"arxiv","id":"2512.20806","version":3},"attestation_state":"computed","paper":{"title":"Safety Alignment of LMs via Non-cooperative Games","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":[],"primary_cat":"cs.AI","authors_text":"Anselm Paulus, Arman Zharmagambetov, Brandon Amos, Ilia Kulikov, Ivan Evtimov, Kamalika Chaudhuri, R\\'emi Munos","submitted_at":"2025-12-23T22:13:14Z","abstract_excerpt":"Ensuring the safety of language models (LMs) while maintaining their usefulness remains a critical challenge in AI alignment. Current approaches rely on sequential adversarial training: generating adversarial prompts and fine-tuning LMs to defend against them. We introduce a different paradigm: framing safety alignment as a non-zero-sum game between an Attacker LM and a Defender LM trained jointly via online reinforcement learning. Each LM continuously adapts to the other's evolving strategies, driving iterative improvement. Our method uses a preference-based reward signal derived from pairwis"},"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":"2512.20806","kind":"arxiv","version":3},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.AI","submitted_at":"2025-12-23T22:13:14Z","cross_cats_sorted":[],"title_canon_sha256":"937ce726892f083bd9013aefa1ddfb1278bbb94d89d7242eea43ad75cf3ee50c","abstract_canon_sha256":"21a3fd2105f892ad71593d5fce318ca63f022c90f7c5d644e830b292a503fa88"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-06-02T02:04:12.670311Z","signature_b64":"qgFhGHlzH7lkJu/4OHLq+A0GWk4jIQDeaIzcdL00zsfKABLtKaYxgLiiCrVgxiPjewH9oiGgGQborCprOxIhBg==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"3cf8be88ba8dccaca7b5d7fb2d875e40c14dde51753bdf4e2f008febe1ca32f0","last_reissued_at":"2026-06-02T02:04:12.669805Z","signature_status":"signed_v1","first_computed_at":"2026-06-02T02:04:12.669805Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Safety Alignment of LMs via Non-cooperative Games","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":[],"primary_cat":"cs.AI","authors_text":"Anselm Paulus, Arman Zharmagambetov, Brandon Amos, Ilia Kulikov, Ivan Evtimov, Kamalika Chaudhuri, R\\'emi Munos","submitted_at":"2025-12-23T22:13:14Z","abstract_excerpt":"Ensuring the safety of language models (LMs) while maintaining their usefulness remains a critical challenge in AI alignment. Current approaches rely on sequential adversarial training: generating adversarial prompts and fine-tuning LMs to defend against them. We introduce a different paradigm: framing safety alignment as a non-zero-sum game between an Attacker LM and a Defender LM trained jointly via online reinforcement learning. Each LM continuously adapts to the other's evolving strategies, driving iterative improvement. Our method uses a preference-based reward signal derived from pairwis"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2512.20806","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/2512.20806/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":"2512.20806","created_at":"2026-06-02T02:04:12.669859+00:00"},{"alias_kind":"arxiv_version","alias_value":"2512.20806v3","created_at":"2026-06-02T02:04:12.669859+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2512.20806","created_at":"2026-06-02T02:04:12.669859+00:00"},{"alias_kind":"pith_short_12","alias_value":"HT4L5CF2RXGK","created_at":"2026-06-02T02:04:12.669859+00:00"},{"alias_kind":"pith_short_16","alias_value":"HT4L5CF2RXGKZJ5V","created_at":"2026-06-02T02:04:12.669859+00:00"},{"alias_kind":"pith_short_8","alias_value":"HT4L5CF2","created_at":"2026-06-02T02:04:12.669859+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":1,"internal_anchor_count":1,"sample":[{"citing_arxiv_id":"2605.13806","citing_title":"Min-Max Optimization Requires Exponentially Many Queries","ref_index":6,"is_internal_anchor":true}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/HT4L5CF2RXGKZJ5V275S3B26ID","json":"https://pith.science/pith/HT4L5CF2RXGKZJ5V275S3B26ID.json","graph_json":"https://pith.science/api/pith-number/HT4L5CF2RXGKZJ5V275S3B26ID/graph.json","events_json":"https://pith.science/api/pith-number/HT4L5CF2RXGKZJ5V275S3B26ID/events.json","paper":"https://pith.science/paper/HT4L5CF2"},"agent_actions":{"view_html":"https://pith.science/pith/HT4L5CF2RXGKZJ5V275S3B26ID","download_json":"https://pith.science/pith/HT4L5CF2RXGKZJ5V275S3B26ID.json","view_paper":"https://pith.science/paper/HT4L5CF2","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2512.20806&json=true","fetch_graph":"https://pith.science/api/pith-number/HT4L5CF2RXGKZJ5V275S3B26ID/graph.json","fetch_events":"https://pith.science/api/pith-number/HT4L5CF2RXGKZJ5V275S3B26ID/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/HT4L5CF2RXGKZJ5V275S3B26ID/action/timestamp_anchor","attest_storage":"https://pith.science/pith/HT4L5CF2RXGKZJ5V275S3B26ID/action/storage_attestation","attest_author":"https://pith.science/pith/HT4L5CF2RXGKZJ5V275S3B26ID/action/author_attestation","sign_citation":"https://pith.science/pith/HT4L5CF2RXGKZJ5V275S3B26ID/action/citation_signature","submit_replication":"https://pith.science/pith/HT4L5CF2RXGKZJ5V275S3B26ID/action/replication_record"}},"created_at":"2026-06-02T02:04:12.669859+00:00","updated_at":"2026-06-02T02:04:12.669859+00:00"}