{"paper":{"title":"Over-Refusal and Representation Subspaces: A Mechanistic Analysis of Task-Conditioned Refusal in Aligned LLMs","license":"http://creativecommons.org/licenses/by/4.0/","headline":"Aligned LLMs encode harmful refusal in one global hidden-state direction but spread over-refusal across separate task-specific subspaces.","cross_cats":[],"primary_cat":"cs.CL","authors_text":"Mark Dras, Usman Naseem, Utsav Maskey","submitted_at":"2026-03-29T04:53:40Z","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"},"claims":{"count":4,"items":[{"kind":"strongest_claim","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.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"That linear probes and direction-finding methods on hidden states accurately isolate the causal mechanisms of refusal rather than merely capturing correlational patterns.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","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.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"Aligned LLMs encode harmful refusal in one global hidden-state direction but spread over-refusal across separate task-specific subspaces.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"b94d0ddb5c9a53735530c111f2ea71b1a68eb1a50e0d1bfa2851fe702c7b5d5e"},"source":{"id":"2603.27518","kind":"arxiv","version":3},"verdict":{"id":"b2a85814-1f52-4a40-a141-35f7580b78e8","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-14T22:15:18.036540Z","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.","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","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.","pith_extraction_headline":"Aligned LLMs encode harmful refusal in one global hidden-state direction but spread over-refusal across separate task-specific subspaces."},"integrity":{"clean":true,"summary":{"advisory":0,"critical":0,"by_detector":{},"informational":0},"endpoint":"/pith/2603.27518/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"}