{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2025:6UZGADZMYHG6NVBEIAIFLUXH7X","short_pith_number":"pith:6UZGADZM","schema_version":"1.0","canonical_sha256":"f532600f2cc1cde6d424401055d2e7fddaa2530b4ebbce14d5729e9f4a077a3a","source":{"kind":"arxiv","id":"2505.23556","version":1},"attestation_state":"computed","paper":{"title":"Understanding Refusal in Language Models with Sparse Autoencoders","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":[],"primary_cat":"cs.CL","authors_text":"Clement Neo, Erik Cambria, Nirmalendu Prakash, Ranjan Satapathy, Roy Ka-Wei Lee, Wei Jie Yeo","submitted_at":"2025-05-29T15:33:39Z","abstract_excerpt":"Refusal is a key safety behavior in aligned language models, yet the internal mechanisms driving refusals remain opaque. In this work, we conduct a mechanistic study of refusal in instruction-tuned LLMs using sparse autoencoders to identify latent features that causally mediate refusal behaviors. We apply our method to two open-source chat models and intervene on refusal-related features to assess their influence on generation, validating their behavioral impact across multiple harmful datasets. This enables a fine-grained inspection of how refusal manifests at the activation level and address"},"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":"2505.23556","kind":"arxiv","version":1},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.CL","submitted_at":"2025-05-29T15:33:39Z","cross_cats_sorted":[],"title_canon_sha256":"9fb52cffbc396e5ae8968e55ce022d4154bc71c629fdd3e0b3f544dead9e2803","abstract_canon_sha256":"f8977356cb7d094d15c1bd682b0c120ac401f76d4709be723c7fc30fadc4acc1"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T11:12:03.862859Z","signature_b64":"cQ4A/DlGLzv42WgjIpAnYlxDp+3IkXqozE3TTrjc7YKlA0O0VEwr+NMctCKtchnEefVId9K2gk6Z/PioOv1lDg==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"f532600f2cc1cde6d424401055d2e7fddaa2530b4ebbce14d5729e9f4a077a3a","last_reissued_at":"2026-07-05T11:12:03.862380Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T11:12:03.862380Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Understanding Refusal in Language Models with Sparse Autoencoders","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":[],"primary_cat":"cs.CL","authors_text":"Clement Neo, Erik Cambria, Nirmalendu Prakash, Ranjan Satapathy, Roy Ka-Wei Lee, Wei Jie Yeo","submitted_at":"2025-05-29T15:33:39Z","abstract_excerpt":"Refusal is a key safety behavior in aligned language models, yet the internal mechanisms driving refusals remain opaque. In this work, we conduct a mechanistic study of refusal in instruction-tuned LLMs using sparse autoencoders to identify latent features that causally mediate refusal behaviors. We apply our method to two open-source chat models and intervene on refusal-related features to assess their influence on generation, validating their behavioral impact across multiple harmful datasets. This enables a fine-grained inspection of how refusal manifests at the activation level and address"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2505.23556","kind":"arxiv","version":1},"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/2505.23556/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":"2505.23556","created_at":"2026-07-05T11:12:03.862433+00:00"},{"alias_kind":"arxiv_version","alias_value":"2505.23556v1","created_at":"2026-07-05T11:12:03.862433+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2505.23556","created_at":"2026-07-05T11:12:03.862433+00:00"},{"alias_kind":"pith_short_12","alias_value":"6UZGADZMYHG6","created_at":"2026-07-05T11:12:03.862433+00:00"},{"alias_kind":"pith_short_16","alias_value":"6UZGADZMYHG6NVBE","created_at":"2026-07-05T11:12:03.862433+00:00"},{"alias_kind":"pith_short_8","alias_value":"6UZGADZM","created_at":"2026-07-05T11:12:03.862433+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":4,"internal_anchor_count":1,"sample":[{"citing_arxiv_id":"2607.05355","citing_title":"Faithfulness to Refusal: A Causal Audit of Neuron Selectors","ref_index":33,"is_internal_anchor":true},{"citing_arxiv_id":"2607.02047","citing_title":"OpenSafeIntent: Evaluating Intent-Calibrated Safe Completion Across Dual-Use Prompt Sets","ref_index":3,"is_internal_anchor":false},{"citing_arxiv_id":"2606.15980","citing_title":"Do Activation Monitors Survive Model Updates? Benchmarking, Predicting, and Repairing Activation-Monitor Staleness","ref_index":45,"is_internal_anchor":false},{"citing_arxiv_id":"2605.07883","citing_title":"Beyond \"I cannot fulfill this request\": Alleviating Rigid Rejection in LLMs via Label Enhancement","ref_index":11,"is_internal_anchor":false}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/6UZGADZMYHG6NVBEIAIFLUXH7X","json":"https://pith.science/pith/6UZGADZMYHG6NVBEIAIFLUXH7X.json","graph_json":"https://pith.science/api/pith-number/6UZGADZMYHG6NVBEIAIFLUXH7X/graph.json","events_json":"https://pith.science/api/pith-number/6UZGADZMYHG6NVBEIAIFLUXH7X/events.json","paper":"https://pith.science/paper/6UZGADZM"},"agent_actions":{"view_html":"https://pith.science/pith/6UZGADZMYHG6NVBEIAIFLUXH7X","download_json":"https://pith.science/pith/6UZGADZMYHG6NVBEIAIFLUXH7X.json","view_paper":"https://pith.science/paper/6UZGADZM","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2505.23556&json=true","fetch_graph":"https://pith.science/api/pith-number/6UZGADZMYHG6NVBEIAIFLUXH7X/graph.json","fetch_events":"https://pith.science/api/pith-number/6UZGADZMYHG6NVBEIAIFLUXH7X/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/6UZGADZMYHG6NVBEIAIFLUXH7X/action/timestamp_anchor","attest_storage":"https://pith.science/pith/6UZGADZMYHG6NVBEIAIFLUXH7X/action/storage_attestation","attest_author":"https://pith.science/pith/6UZGADZMYHG6NVBEIAIFLUXH7X/action/author_attestation","sign_citation":"https://pith.science/pith/6UZGADZMYHG6NVBEIAIFLUXH7X/action/citation_signature","submit_replication":"https://pith.science/pith/6UZGADZMYHG6NVBEIAIFLUXH7X/action/replication_record"}},"created_at":"2026-07-05T11:12:03.862433+00:00","updated_at":"2026-07-05T11:12:03.862433+00:00"}