{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2026:3EUUJX67Z5S2ATHNLCN7UPXCJA","short_pith_number":"pith:3EUUJX67","schema_version":"1.0","canonical_sha256":"d92944dfdfcf65a04ced589bfa3ee2481bd135a94f0d604ef1c05e5bafefb3ca","source":{"kind":"arxiv","id":"2604.27019","version":2},"attestation_state":"computed","paper":{"title":"Dynamic Adversarial Fine-Tuning Reorganizes Refusal Geometry","license":"http://creativecommons.org/licenses/by-nc-nd/4.0/","headline":"Dynamic adversarial fine-tuning reorganizes refusal geometry by relocating the primary carrier from late layers to early layers.","cross_cats":["cs.CL","cs.CR"],"primary_cat":"cs.LG","authors_text":"Haihua Shen, Junbin Yang, Shan Li, Wenhao Lan, Yijun Yang","submitted_at":"2026-04-29T12:44:05Z","abstract_excerpt":"Safety-aligned language models must refuse harmful requests without collapsing into broad over-refusal, yet it remains unclear how dynamic adversarial fine-tuning changes the internal carriers of refusal. We study one 7B backbone under supervised fine-tuning (SFT) and under Robust Refusal Dynamic Defense (R2D2), a HarmBench-style adversarial fine-tuning procedure that repeatedly refreshes harmful training cases with current jailbreak attacks. Our protocol aligns fixed-source HarmBench, StrongREJECT, and XSTest with a five-anchor refusal-geometry suite, causal interventions, and a sparse adapti"},"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":"2604.27019","kind":"arxiv","version":2},"metadata":{"license":"http://creativecommons.org/licenses/by-nc-nd/4.0/","primary_cat":"cs.LG","submitted_at":"2026-04-29T12:44:05Z","cross_cats_sorted":["cs.CL","cs.CR"],"title_canon_sha256":"927b03b1fcd80ef19340f467342c19c575ccadfebac359d713a90831292e5fed","abstract_canon_sha256":"a33e5c77f8972c6899dffa859d630a2f0e07185da66e708ef465b7cc16a55450"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-05-20T00:03:13.065351Z","signature_b64":"uhozzIjoPL0pU7A3a6rNsQWtMgqYXTiG5R78hl0nVzjx7DWAdxBGvaYu5vmKTl0rSL/et0akBoVOxBb3kZ6IBQ==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"d92944dfdfcf65a04ced589bfa3ee2481bd135a94f0d604ef1c05e5bafefb3ca","last_reissued_at":"2026-05-20T00:03:13.064371Z","signature_status":"signed_v1","first_computed_at":"2026-05-20T00:03:13.064371Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Dynamic Adversarial Fine-Tuning Reorganizes Refusal Geometry","license":"http://creativecommons.org/licenses/by-nc-nd/4.0/","headline":"Dynamic adversarial fine-tuning reorganizes refusal geometry by relocating the primary carrier from late layers to early layers.","cross_cats":["cs.CL","cs.CR"],"primary_cat":"cs.LG","authors_text":"Haihua Shen, Junbin Yang, Shan Li, Wenhao Lan, Yijun Yang","submitted_at":"2026-04-29T12:44:05Z","abstract_excerpt":"Safety-aligned language models must refuse harmful requests without collapsing into broad over-refusal, yet it remains unclear how dynamic adversarial fine-tuning changes the internal carriers of refusal. We study one 7B backbone under supervised fine-tuning (SFT) and under Robust Refusal Dynamic Defense (R2D2), a HarmBench-style adversarial fine-tuning procedure that repeatedly refreshes harmful training cases with current jailbreak attacks. Our protocol aligns fixed-source HarmBench, StrongREJECT, and XSTest with a five-anchor refusal-geometry suite, causal interventions, and a sparse adapti"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"These results support a reorganization account rather than a drift-only account, with evidence limited to one backbone and fixed-source attacks. R2D2 preserves a late-layer admissible carrier through step 100 before relocating to an early-layer carrier, while effective rank remains near 1.23--1.27.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"That the five-anchor refusal-geometry suite and causal interventions accurately isolate reorganization effects without being confounded by the specific 7B backbone, fixed attack sources, or unmeasured factors in the training dynamics.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"R2D2-style dynamic adversarial fine-tuning reorganizes refusal geometry from late-layer to early-layer carriers in LLMs, achieving lower attack success rates than SFT while maintaining low-dimensional structure.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"Dynamic adversarial fine-tuning reorganizes refusal geometry by relocating the primary carrier from late layers to early layers.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"265c253502e11c0992554407c8dbb6ac506b4fd15e594aafd951c423a0df48b5"},"source":{"id":"2604.27019","kind":"arxiv","version":2},"verdict":{"id":"93cead99-4432-44c0-89a2-9ac4e8473454","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-07T11:32:46.163951Z","strongest_claim":"These results support a reorganization account rather than a drift-only account, with evidence limited to one backbone and fixed-source attacks. R2D2 preserves a late-layer admissible carrier through step 100 before relocating to an early-layer carrier, while effective rank remains near 1.23--1.27.","one_line_summary":"R2D2-style dynamic adversarial fine-tuning reorganizes refusal geometry from late-layer to early-layer carriers in LLMs, achieving lower attack success rates than SFT while maintaining low-dimensional structure.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"That the five-anchor refusal-geometry suite and causal interventions accurately isolate reorganization effects without being confounded by the specific 7B backbone, fixed attack sources, or unmeasured factors in the training dynamics.","pith_extraction_headline":"Dynamic adversarial fine-tuning reorganizes refusal geometry by relocating the primary carrier from late layers to early layers."},"integrity":{"clean":true,"summary":{"advisory":0,"critical":0,"by_detector":{},"informational":0},"endpoint":"/pith/2604.27019/integrity.json","findings":[],"available":true,"detectors_run":[{"name":"doi_compliance","ran_at":"2026-05-19T19:58:34.044956Z","status":"completed","version":"1.0.0","findings_count":0}],"snapshot_sha256":"5847dc47980c3abd0858840420a67c44eebbfd69949c723ee090881a0ba9d545"},"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":"2604.27019","created_at":"2026-05-20T00:03:13.064520+00:00"},{"alias_kind":"arxiv_version","alias_value":"2604.27019v2","created_at":"2026-05-20T00:03:13.064520+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2604.27019","created_at":"2026-05-20T00:03:13.064520+00:00"},{"alias_kind":"pith_short_12","alias_value":"3EUUJX67Z5S2","created_at":"2026-05-20T00:03:13.064520+00:00"},{"alias_kind":"pith_short_16","alias_value":"3EUUJX67Z5S2ATHN","created_at":"2026-05-20T00:03:13.064520+00:00"},{"alias_kind":"pith_short_8","alias_value":"3EUUJX67","created_at":"2026-05-20T00:03:13.064520+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":0,"internal_anchor_count":0,"sample":[]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/3EUUJX67Z5S2ATHNLCN7UPXCJA","json":"https://pith.science/pith/3EUUJX67Z5S2ATHNLCN7UPXCJA.json","graph_json":"https://pith.science/api/pith-number/3EUUJX67Z5S2ATHNLCN7UPXCJA/graph.json","events_json":"https://pith.science/api/pith-number/3EUUJX67Z5S2ATHNLCN7UPXCJA/events.json","paper":"https://pith.science/paper/3EUUJX67"},"agent_actions":{"view_html":"https://pith.science/pith/3EUUJX67Z5S2ATHNLCN7UPXCJA","download_json":"https://pith.science/pith/3EUUJX67Z5S2ATHNLCN7UPXCJA.json","view_paper":"https://pith.science/paper/3EUUJX67","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2604.27019&json=true","fetch_graph":"https://pith.science/api/pith-number/3EUUJX67Z5S2ATHNLCN7UPXCJA/graph.json","fetch_events":"https://pith.science/api/pith-number/3EUUJX67Z5S2ATHNLCN7UPXCJA/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/3EUUJX67Z5S2ATHNLCN7UPXCJA/action/timestamp_anchor","attest_storage":"https://pith.science/pith/3EUUJX67Z5S2ATHNLCN7UPXCJA/action/storage_attestation","attest_author":"https://pith.science/pith/3EUUJX67Z5S2ATHNLCN7UPXCJA/action/author_attestation","sign_citation":"https://pith.science/pith/3EUUJX67Z5S2ATHNLCN7UPXCJA/action/citation_signature","submit_replication":"https://pith.science/pith/3EUUJX67Z5S2ATHNLCN7UPXCJA/action/replication_record"}},"created_at":"2026-05-20T00:03:13.064520+00:00","updated_at":"2026-05-20T00:03:13.064520+00:00"}