{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2025:OVYMK562XE2E4WR2SW3GVIF2O2","short_pith_number":"pith:OVYMK562","schema_version":"1.0","canonical_sha256":"7570c577dab9344e5a3a95b66aa0ba76a14a19d32a69e1157f0d1a599f91ba19","source":{"kind":"arxiv","id":"2507.11878","version":5},"attestation_state":"computed","paper":{"title":"LLMs Encode Harmfulness and Refusal Separately","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":[],"primary_cat":"cs.CL","authors_text":"David Bau, Jiachen Zhao, Jing Huang, Weiyan Shi, Zhengxuan Wu","submitted_at":"2025-07-16T03:48:03Z","abstract_excerpt":"LLMs are trained to refuse harmful instructions, but do they truly understand harmfulness beyond just refusing? Prior work has shown that LLMs' refusal behaviors can be mediated by a one-dimensional subspace, i.e., a refusal direction. In this work, we identify a new dimension to analyze safety mechanisms in LLMs, i.e., harmfulness, which is encoded internally as a separate concept from refusal. There exists a harmfulness direction that is distinct from the refusal direction. As causal evidence, steering along the harmfulness direction can lead LLMs to interpret harmless instructions as harmfu"},"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":"2507.11878","kind":"arxiv","version":5},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.CL","submitted_at":"2025-07-16T03:48:03Z","cross_cats_sorted":[],"title_canon_sha256":"ec01ad32a6ebee1860a08defe268f753fe7244ef5bbcade7b3d4bb663cc3d1e9","abstract_canon_sha256":"811fa6326ab2e562a174f72e5dd0e5825857b4e4dd4d63e546b42d1664c08e0b"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-07T02:18:25.901498Z","signature_b64":"x6CiWCNKII6hRZ/Ukrs/4AZRBbVw9xcOAGxUprlzUujhHFIz69y4cehigzcH9JVVsNyQxfMh3bKJR4YQqN5HBw==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"7570c577dab9344e5a3a95b66aa0ba76a14a19d32a69e1157f0d1a599f91ba19","last_reissued_at":"2026-07-07T02:18:25.900807Z","signature_status":"signed_v1","first_computed_at":"2026-07-07T02:18:25.900807Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"LLMs Encode Harmfulness and Refusal Separately","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":[],"primary_cat":"cs.CL","authors_text":"David Bau, Jiachen Zhao, Jing Huang, Weiyan Shi, Zhengxuan Wu","submitted_at":"2025-07-16T03:48:03Z","abstract_excerpt":"LLMs are trained to refuse harmful instructions, but do they truly understand harmfulness beyond just refusing? Prior work has shown that LLMs' refusal behaviors can be mediated by a one-dimensional subspace, i.e., a refusal direction. In this work, we identify a new dimension to analyze safety mechanisms in LLMs, i.e., harmfulness, which is encoded internally as a separate concept from refusal. There exists a harmfulness direction that is distinct from the refusal direction. As causal evidence, steering along the harmfulness direction can lead LLMs to interpret harmless instructions as harmfu"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2507.11878","kind":"arxiv","version":5},"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/2507.11878/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":"2507.11878","created_at":"2026-07-07T02:18:25.900910+00:00"},{"alias_kind":"arxiv_version","alias_value":"2507.11878v5","created_at":"2026-07-07T02:18:25.900910+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2507.11878","created_at":"2026-07-07T02:18:25.900910+00:00"},{"alias_kind":"pith_short_12","alias_value":"OVYMK562XE2E","created_at":"2026-07-07T02:18:25.900910+00:00"},{"alias_kind":"pith_short_16","alias_value":"OVYMK562XE2E4WR2","created_at":"2026-07-07T02:18:25.900910+00:00"},{"alias_kind":"pith_short_8","alias_value":"OVYMK562","created_at":"2026-07-07T02:18:25.900910+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":16,"internal_anchor_count":16,"sample":[{"citing_arxiv_id":"2607.02047","citing_title":"OpenSafeIntent: Evaluating Intent-Calibrated Safe Completion Across Dual-Use Prompt Sets","ref_index":1,"is_internal_anchor":true},{"citing_arxiv_id":"2606.08044","citing_title":"When Behavioral Safety Evaluation Fails: A Representation-Level Perspective","ref_index":11,"is_internal_anchor":true},{"citing_arxiv_id":"2606.07335","citing_title":"Defending Jailbreak Attacks on Large Language Models via Manifold Trajectory Kinetics","ref_index":63,"is_internal_anchor":true},{"citing_arxiv_id":"2606.06667","citing_title":"The Piggyback Hypothesis of Generalization: Explaining and Mitigating Emergent Misalignment","ref_index":29,"is_internal_anchor":true},{"citing_arxiv_id":"2606.01196","citing_title":"Low-Resource Safety Failures Are Action Failures, Not Representation Failures","ref_index":31,"is_internal_anchor":true},{"citing_arxiv_id":"2605.13339","citing_title":"Probing Persona-Dependent Preferences in Language Models","ref_index":34,"is_internal_anchor":true},{"citing_arxiv_id":"2605.14038","citing_title":"Model-Adaptive Tool Necessity Reveals the Knowing-Doing Gap in LLM Tool Use","ref_index":37,"is_internal_anchor":true},{"citing_arxiv_id":"2510.24941","citing_title":"Can Aha Moments Be Fake? Towards Quantifying Decorative and True Thinking in Chain-of-Thought","ref_index":36,"is_internal_anchor":true},{"citing_arxiv_id":"2605.14038","citing_title":"Model-Adaptive Tool Necessity Reveals the Knowing-Doing Gap in LLM Tool Use","ref_index":37,"is_internal_anchor":true},{"citing_arxiv_id":"2605.12726","citing_title":"Before the Last Token: Diagnosing Final-Token Safety Probe Failures","ref_index":14,"is_internal_anchor":true},{"citing_arxiv_id":"2604.06233","citing_title":"Blind Refusal: Language Models Refuse to Help Users Evade Unjust, Absurd, and Illegitimate Rules","ref_index":38,"is_internal_anchor":true},{"citing_arxiv_id":"2604.18901","citing_title":"Harmful Intent as a Geometrically Recoverable Feature of LLM Residual Streams","ref_index":13,"is_internal_anchor":true},{"citing_arxiv_id":"2604.18510","citing_title":"Different Paths to Harmful Compliance: Behavioral Side Effects and Mechanistic Divergence Across LLM Jailbreaks","ref_index":40,"is_internal_anchor":true},{"citing_arxiv_id":"2604.04385","citing_title":"How Alignment Routes: Localizing, Scaling, and Controlling Policy Circuits in Language Models","ref_index":10,"is_internal_anchor":true},{"citing_arxiv_id":"2604.18519","citing_title":"LLM Safety From Within: Detecting Harmful Content with Internal Representations","ref_index":11,"is_internal_anchor":true},{"citing_arxiv_id":"2604.18901","citing_title":"Harmful Intent as a Geometrically Recoverable Feature of LLM Residual Streams","ref_index":13,"is_internal_anchor":true}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/OVYMK562XE2E4WR2SW3GVIF2O2","json":"https://pith.science/pith/OVYMK562XE2E4WR2SW3GVIF2O2.json","graph_json":"https://pith.science/api/pith-number/OVYMK562XE2E4WR2SW3GVIF2O2/graph.json","events_json":"https://pith.science/api/pith-number/OVYMK562XE2E4WR2SW3GVIF2O2/events.json","paper":"https://pith.science/paper/OVYMK562"},"agent_actions":{"view_html":"https://pith.science/pith/OVYMK562XE2E4WR2SW3GVIF2O2","download_json":"https://pith.science/pith/OVYMK562XE2E4WR2SW3GVIF2O2.json","view_paper":"https://pith.science/paper/OVYMK562","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2507.11878&json=true","fetch_graph":"https://pith.science/api/pith-number/OVYMK562XE2E4WR2SW3GVIF2O2/graph.json","fetch_events":"https://pith.science/api/pith-number/OVYMK562XE2E4WR2SW3GVIF2O2/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/OVYMK562XE2E4WR2SW3GVIF2O2/action/timestamp_anchor","attest_storage":"https://pith.science/pith/OVYMK562XE2E4WR2SW3GVIF2O2/action/storage_attestation","attest_author":"https://pith.science/pith/OVYMK562XE2E4WR2SW3GVIF2O2/action/author_attestation","sign_citation":"https://pith.science/pith/OVYMK562XE2E4WR2SW3GVIF2O2/action/citation_signature","submit_replication":"https://pith.science/pith/OVYMK562XE2E4WR2SW3GVIF2O2/action/replication_record"}},"created_at":"2026-07-07T02:18:25.900910+00:00","updated_at":"2026-07-07T02:18:25.900910+00:00"}