{"state_type":"pith_open_graph_state","state_version":"1.0","pith_number":"pith:2026:SUDNVRJXYPEWMY7LUGANHCHAKU","merge_version":"pith-open-graph-merge-v1","event_count":2,"valid_event_count":2,"invalid_event_count":0,"equivocation_count":0,"current":{"canonical_record":{"metadata":{"abstract_canon_sha256":"36a047054b045a16488b2793b3efb48cac8af020700fbefe2d1ca71f664e5fb3","cross_cats_sorted":[],"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.CL","submitted_at":"2026-03-29T04:53:40Z","title_canon_sha256":"b7dd9d502d6f0c0f6a781fdf84e1f7fcd69d73c91989503844322f1f8955580e"},"schema_version":"1.0","source":{"id":"2603.27518","kind":"arxiv","version":3}},"source_aliases":[{"alias_kind":"arxiv","alias_value":"2603.27518","created_at":"2026-05-29T01:05:08Z"},{"alias_kind":"arxiv_version","alias_value":"2603.27518v3","created_at":"2026-05-29T01:05:08Z"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2603.27518","created_at":"2026-05-29T01:05:08Z"},{"alias_kind":"pith_short_12","alias_value":"SUDNVRJXYPEW","created_at":"2026-05-29T01:05:08Z"},{"alias_kind":"pith_short_16","alias_value":"SUDNVRJXYPEWMY7L","created_at":"2026-05-29T01:05:08Z"},{"alias_kind":"pith_short_8","alias_value":"SUDNVRJX","created_at":"2026-05-29T01:05:08Z"}],"graph_snapshots":[{"event_id":"sha256:3832e3a1d97b9f247b6f42c866fb3db3a77f85f71eb45bae9307ab258ac3e133","target":"graph","created_at":"2026-05-29T01:05:08Z","signer":{"key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54","signer_id":"pith.science","signer_type":"pith_registry"},"payload":{"graph_snapshot":{"author_claims":{"count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57","strong_count":0},"builder_version":"pith-number-builder-2026-05-17-v1","claims":{"count":4,"items":[{"attestation":"unclaimed","claim_id":"C1","kind":"strongest_claim","source":"verdict.strongest_claim","status":"machine_extracted","text":"harmful-refusal directions are task-agnostic and can be captured by a single global vector, whereas over-refusal directions are task-dependent: they reside within the benign task-representation clusters, vary across tasks, and span a higher-dimensional subspace."},{"attestation":"unclaimed","claim_id":"C2","kind":"weakest_assumption","source":"verdict.weakest_assumption","status":"machine_extracted","text":"That linear probes and direction-finding methods on hidden states accurately isolate the causal mechanisms of refusal rather than merely capturing correlational patterns."},{"attestation":"unclaimed","claim_id":"C3","kind":"one_line_summary","source":"verdict.one_line_summary","status":"machine_extracted","text":"Harmful refusal in aligned LLMs is captured by a single task-agnostic vector, but over-refusal directions are task-dependent, reside in benign task clusters, and span higher-dimensional subspaces."},{"attestation":"unclaimed","claim_id":"C4","kind":"headline","source":"verdict.pith_extraction.headline","status":"machine_extracted","text":"Aligned LLMs encode harmful refusal in one global hidden-state direction but spread over-refusal across separate task-specific subspaces."}],"snapshot_sha256":"b94d0ddb5c9a53735530c111f2ea71b1a68eb1a50e0d1bfa2851fe702c7b5d5e"},"formal_canon":{"evidence_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"integrity":{"available":true,"clean":true,"detectors_run":[],"endpoint":"/pith/2603.27518/integrity.json","findings":[],"snapshot_sha256":"c28c3603d3b5d939e8dc4c7e95fa8dfce3d595e45f758748cecf8e644a296938","summary":{"advisory":0,"by_detector":{},"critical":0,"informational":0}},"paper":{"abstract_excerpt":"Aligned language models that are trained to refuse harmful requests also exhibit over-refusal: they decline safe instructions that seemingly resemble harmful instructions. A natural approach is to ablate the global refusal direction, steering the hidden-state vectors away or towards the harmful-refusal examples, but this corrects over-refusal only incidentally while disrupting the broader refusal mechanism. In this work, we analyse the representational geometry of both refusal types to understand why this happens. We show that harmful-refusal directions are task-agnostic and can be captured by","authors_text":"Mark Dras, Usman Naseem, Utsav Maskey","cross_cats":[],"headline":"Aligned LLMs encode harmful refusal in one global hidden-state direction but spread over-refusal across separate task-specific subspaces.","license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.CL","submitted_at":"2026-03-29T04:53:40Z","title":"Over-Refusal and Representation Subspaces: A Mechanistic Analysis of Task-Conditioned Refusal in Aligned LLMs"},"references":{"count":0,"internal_anchors":0,"resolved_work":0,"sample":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2603.27518","kind":"arxiv","version":3},"verdict":{"created_at":"2026-05-14T22:15:18.036540Z","id":"b2a85814-1f52-4a40-a141-35f7580b78e8","model_set":{"reader":"grok-4.3"},"one_line_summary":"Harmful refusal in aligned LLMs is captured by a single task-agnostic vector, but over-refusal directions are task-dependent, reside in benign task clusters, and span higher-dimensional subspaces.","pipeline_version":"pith-pipeline@v0.9.0","pith_extraction_headline":"Aligned LLMs encode harmful refusal in one global hidden-state direction but spread over-refusal across separate task-specific subspaces.","strongest_claim":"harmful-refusal directions are task-agnostic and can be captured by a single global vector, whereas over-refusal directions are task-dependent: they reside within the benign task-representation clusters, vary across tasks, and span a higher-dimensional subspace.","weakest_assumption":"That linear probes and direction-finding methods on hidden states accurately isolate the causal mechanisms of refusal rather than merely capturing correlational patterns."}},"verdict_id":"b2a85814-1f52-4a40-a141-35f7580b78e8"}}],"author_attestations":[],"timestamp_anchors":[],"storage_attestations":[],"citation_signatures":[],"replication_records":[],"corrections":[],"mirror_hints":[],"record_created":{"event_id":"sha256:d0b7633b3aac838c510bbb68df9d2321f39c052b9605dd95ed02532a2e370068","target":"record","created_at":"2026-05-29T01:05:08Z","signer":{"key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54","signer_id":"pith.science","signer_type":"pith_registry"},"payload":{"attestation_state":"computed","canonical_record":{"metadata":{"abstract_canon_sha256":"36a047054b045a16488b2793b3efb48cac8af020700fbefe2d1ca71f664e5fb3","cross_cats_sorted":[],"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.CL","submitted_at":"2026-03-29T04:53:40Z","title_canon_sha256":"b7dd9d502d6f0c0f6a781fdf84e1f7fcd69d73c91989503844322f1f8955580e"},"schema_version":"1.0","source":{"id":"2603.27518","kind":"arxiv","version":3}},"canonical_sha256":"9506dac537c3c96663eba180d388e05530cf4304a998ff07bc025c7c5e6922db","receipt":{"algorithm":"ed25519","builder_version":"pith-number-builder-2026-05-17-v1","canonical_sha256":"9506dac537c3c96663eba180d388e05530cf4304a998ff07bc025c7c5e6922db","first_computed_at":"2026-05-29T01:05:08.264221Z","key_id":"pith-v1-2026-05","kind":"pith_receipt","last_reissued_at":"2026-05-29T01:05:08.264221Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54","receipt_version":"0.3","signature_b64":"OxU0numwt5Et+YRByUHDIrj3YwLRDBZcBFRMuRM6UnO3ZtFe0qCJ4ppMvMiSoLjqq482MZiuWpsXZ42Mx26NDw==","signature_status":"signed_v1","signed_at":"2026-05-29T01:05:08.264951Z","signed_message":"canonical_sha256_bytes"},"source_id":"2603.27518","source_kind":"arxiv","source_version":3}}},"equivocations":[],"invalid_events":[],"applied_event_ids":["sha256:d0b7633b3aac838c510bbb68df9d2321f39c052b9605dd95ed02532a2e370068","sha256:3832e3a1d97b9f247b6f42c866fb3db3a77f85f71eb45bae9307ab258ac3e133"],"state_sha256":"66fd67b8eaa3bae0b87b7780d8e9a9069de642c7f7aae86303093a34d1bdaec0"}