{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2024:CH2BG5A2FQN6GR74Z3AE3IMZRL","short_pith_number":"pith:CH2BG5A2","schema_version":"1.0","canonical_sha256":"11f413741a2c1be347fccec04da1998ac62b49c8d77c05a13fecee0843565416","source":{"kind":"arxiv","id":"2411.11296","version":2},"attestation_state":"computed","paper":{"title":"Steering Language Model Refusal with Sparse Autoencoders","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":[],"primary_cat":"cs.LG","authors_text":"Blake Bullwinkel, David Majercak, Dean Carignan, Eric Horvitz, Forough Poursabzi-Sangdeh, Harsha Nori, Jingya Chen, Kyle O'Brien, Richard Edgar, Xavier Fernandes","submitted_at":"2024-11-18T05:47:02Z","abstract_excerpt":"Responsible deployment of language models requires mechanisms for refusing unsafe prompts while preserving model performance. While most approaches modify model weights through additional training, we explore an alternative: steering model activations at inference time via amplifying sparse autoencoder (SAE) features that mediate refusal. This work uncovers a fundamental tension between SAE steering-based safety improvements and general model capabilities. While feature steering successfully improves robustness against both single-turn and challenging multi-turn jailbreak attempts, we discover"},"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":"2411.11296","kind":"arxiv","version":2},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.LG","submitted_at":"2024-11-18T05:47:02Z","cross_cats_sorted":[],"title_canon_sha256":"f6a5f1fa0958319e821d1d864492b8e2ee8d74fe924f148c61effc00405b4435","abstract_canon_sha256":"ce59db6e76956b420d65cef75a44c38d490b7ed6d91ede8204652699cbfcb5fc"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T11:35:30.824044Z","signature_b64":"ViuGKXRP7OLtzBZCD3JbTIFMrbCZ9tlzh6+tbWCnviCyrU0l3dItexgP9gdrUi/XD28zGl7u1fJ0z9l/TEhVAQ==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"11f413741a2c1be347fccec04da1998ac62b49c8d77c05a13fecee0843565416","last_reissued_at":"2026-07-05T11:35:30.823505Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T11:35:30.823505Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Steering Language Model Refusal with Sparse Autoencoders","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":[],"primary_cat":"cs.LG","authors_text":"Blake Bullwinkel, David Majercak, Dean Carignan, Eric Horvitz, Forough Poursabzi-Sangdeh, Harsha Nori, Jingya Chen, Kyle O'Brien, Richard Edgar, Xavier Fernandes","submitted_at":"2024-11-18T05:47:02Z","abstract_excerpt":"Responsible deployment of language models requires mechanisms for refusing unsafe prompts while preserving model performance. While most approaches modify model weights through additional training, we explore an alternative: steering model activations at inference time via amplifying sparse autoencoder (SAE) features that mediate refusal. This work uncovers a fundamental tension between SAE steering-based safety improvements and general model capabilities. While feature steering successfully improves robustness against both single-turn and challenging multi-turn jailbreak attempts, we discover"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2411.11296","kind":"arxiv","version":2},"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/2411.11296/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":"2411.11296","created_at":"2026-07-05T11:35:30.823561+00:00"},{"alias_kind":"arxiv_version","alias_value":"2411.11296v2","created_at":"2026-07-05T11:35:30.823561+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2411.11296","created_at":"2026-07-05T11:35:30.823561+00:00"},{"alias_kind":"pith_short_12","alias_value":"CH2BG5A2FQN6","created_at":"2026-07-05T11:35:30.823561+00:00"},{"alias_kind":"pith_short_16","alias_value":"CH2BG5A2FQN6GR74","created_at":"2026-07-05T11:35:30.823561+00:00"},{"alias_kind":"pith_short_8","alias_value":"CH2BG5A2","created_at":"2026-07-05T11:35:30.823561+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":15,"internal_anchor_count":0,"sample":[{"citing_arxiv_id":"2606.27199","citing_title":"Forecasting With LLMs: Improved Generalization Through Feature Steering","ref_index":34,"is_internal_anchor":false},{"citing_arxiv_id":"2606.08365","citing_title":"Pre-Intervention Prediction of Sparse Autoencoder Steering Side Effects","ref_index":9,"is_internal_anchor":false},{"citing_arxiv_id":"2606.03002","citing_title":"Perplexity Can Miss SAE Feature Damage Under Quantization","ref_index":13,"is_internal_anchor":false},{"citing_arxiv_id":"2605.24154","citing_title":"Palette: A Modular, Controllable, and Efficient Framework for On-demand Authorized Safety Alignment Relaxation in LLMs","ref_index":62,"is_internal_anchor":false},{"citing_arxiv_id":"2606.26396","citing_title":"At the Edge of Understanding: Sparse Autoencoders Trace The Limits of Transformer Generalization","ref_index":37,"is_internal_anchor":false},{"citing_arxiv_id":"2605.23040","citing_title":"Steered Generation via Gradient-Based Optimization on Sparse Query Features","ref_index":32,"is_internal_anchor":false},{"citing_arxiv_id":"2605.16362","citing_title":"When Is Rank-1 Steering Cheap? Geometry, Granularity, and Budgeted Search","ref_index":22,"is_internal_anchor":false},{"citing_arxiv_id":"2605.16362","citing_title":"When Is Rank-1 Steering Cheap? Geometry, Granularity, and Budgeted Search","ref_index":22,"is_internal_anchor":false},{"citing_arxiv_id":"2605.17413","citing_title":"Ablating Safety: Mechanisms for Removing Alignment in Language Models for Security Applications","ref_index":25,"is_internal_anchor":false},{"citing_arxiv_id":"2509.09708","citing_title":"Beyond I'm Sorry, I Can't: Dissecting Large Language Model Refusal","ref_index":21,"is_internal_anchor":false},{"citing_arxiv_id":"2605.08326","citing_title":"LLM Advertisement based on Neuron Auctions","ref_index":17,"is_internal_anchor":false},{"citing_arxiv_id":"2605.06342","citing_title":"Don't Lose Focus: Activation Steering via Key-Orthogonal Projections","ref_index":24,"is_internal_anchor":false},{"citing_arxiv_id":"2604.22167","citing_title":"Estimating Tail Risks in Language Model Output Distributions","ref_index":30,"is_internal_anchor":false},{"citing_arxiv_id":"2604.08846","citing_title":"Dictionary-Aligned Concept Control for Safeguarding Multimodal LLMs","ref_index":72,"is_internal_anchor":false},{"citing_arxiv_id":"2604.18756","citing_title":"Towards Understanding the Robustness of Sparse Autoencoders","ref_index":22,"is_internal_anchor":false}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/CH2BG5A2FQN6GR74Z3AE3IMZRL","json":"https://pith.science/pith/CH2BG5A2FQN6GR74Z3AE3IMZRL.json","graph_json":"https://pith.science/api/pith-number/CH2BG5A2FQN6GR74Z3AE3IMZRL/graph.json","events_json":"https://pith.science/api/pith-number/CH2BG5A2FQN6GR74Z3AE3IMZRL/events.json","paper":"https://pith.science/paper/CH2BG5A2"},"agent_actions":{"view_html":"https://pith.science/pith/CH2BG5A2FQN6GR74Z3AE3IMZRL","download_json":"https://pith.science/pith/CH2BG5A2FQN6GR74Z3AE3IMZRL.json","view_paper":"https://pith.science/paper/CH2BG5A2","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2411.11296&json=true","fetch_graph":"https://pith.science/api/pith-number/CH2BG5A2FQN6GR74Z3AE3IMZRL/graph.json","fetch_events":"https://pith.science/api/pith-number/CH2BG5A2FQN6GR74Z3AE3IMZRL/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/CH2BG5A2FQN6GR74Z3AE3IMZRL/action/timestamp_anchor","attest_storage":"https://pith.science/pith/CH2BG5A2FQN6GR74Z3AE3IMZRL/action/storage_attestation","attest_author":"https://pith.science/pith/CH2BG5A2FQN6GR74Z3AE3IMZRL/action/author_attestation","sign_citation":"https://pith.science/pith/CH2BG5A2FQN6GR74Z3AE3IMZRL/action/citation_signature","submit_replication":"https://pith.science/pith/CH2BG5A2FQN6GR74Z3AE3IMZRL/action/replication_record"}},"created_at":"2026-07-05T11:35:30.823561+00:00","updated_at":"2026-07-05T11:35:30.823561+00:00"}