{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2024:ZF2G7UWTGDVOZLPQ7DYFZ5X2NJ","short_pith_number":"pith:ZF2G7UWT","schema_version":"1.0","canonical_sha256":"c9746fd2d330eaecadf0f8f05cf6fa6a6af27563a1d346252ad48cca478ebc14","source":{"kind":"arxiv","id":"2403.00745","version":1},"attestation_state":"computed","paper":{"title":"AtP*: An efficient and scalable method for localizing LLM behaviour to components","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.CL"],"primary_cat":"cs.LG","authors_text":"J\\'anos Kram\\'ar, Neel Nanda (Google DeepMind), Rohin Shah, Tom Lieberum","submitted_at":"2024-03-01T18:43:51Z","abstract_excerpt":"Activation Patching is a method of directly computing causal attributions of behavior to model components. However, applying it exhaustively requires a sweep with cost scaling linearly in the number of model components, which can be prohibitively expensive for SoTA Large Language Models (LLMs). We investigate Attribution Patching (AtP), a fast gradient-based approximation to Activation Patching and find two classes of failure modes of AtP which lead to significant false negatives. We propose a variant of AtP called AtP*, with two changes to address these failure modes while retaining scalabili"},"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":"2403.00745","kind":"arxiv","version":1},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.LG","submitted_at":"2024-03-01T18:43:51Z","cross_cats_sorted":["cs.CL"],"title_canon_sha256":"017fb8d34720c998d10a401849eddd0a2c04fbf7d164f73649a52b790c54045f","abstract_canon_sha256":"c5d13ccbf5cf356e9de14b6705ca56b74f12391cea83fc5b247c6a691c1f6b78"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T07:51:08.086285Z","signature_b64":"0/p6ZBqtfnVBc35LHVbB7Wgqpt3FojHpR1kuNwOS3BYMM6peAtLYborPhJq8r+ae9pwKIRwkLef2C+E5tl5MDQ==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"c9746fd2d330eaecadf0f8f05cf6fa6a6af27563a1d346252ad48cca478ebc14","last_reissued_at":"2026-07-05T07:51:08.085735Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T07:51:08.085735Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"AtP*: An efficient and scalable method for localizing LLM behaviour to components","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.CL"],"primary_cat":"cs.LG","authors_text":"J\\'anos Kram\\'ar, Neel Nanda (Google DeepMind), Rohin Shah, Tom Lieberum","submitted_at":"2024-03-01T18:43:51Z","abstract_excerpt":"Activation Patching is a method of directly computing causal attributions of behavior to model components. However, applying it exhaustively requires a sweep with cost scaling linearly in the number of model components, which can be prohibitively expensive for SoTA Large Language Models (LLMs). We investigate Attribution Patching (AtP), a fast gradient-based approximation to Activation Patching and find two classes of failure modes of AtP which lead to significant false negatives. We propose a variant of AtP called AtP*, with two changes to address these failure modes while retaining scalabili"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2403.00745","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/2403.00745/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":"2403.00745","created_at":"2026-07-05T07:51:08.085804+00:00"},{"alias_kind":"arxiv_version","alias_value":"2403.00745v1","created_at":"2026-07-05T07:51:08.085804+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2403.00745","created_at":"2026-07-05T07:51:08.085804+00:00"},{"alias_kind":"pith_short_12","alias_value":"ZF2G7UWTGDVO","created_at":"2026-07-05T07:51:08.085804+00:00"},{"alias_kind":"pith_short_16","alias_value":"ZF2G7UWTGDVOZLPQ","created_at":"2026-07-05T07:51:08.085804+00:00"},{"alias_kind":"pith_short_8","alias_value":"ZF2G7UWT","created_at":"2026-07-05T07:51:08.085804+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":19,"internal_anchor_count":1,"sample":[{"citing_arxiv_id":"2607.08349","citing_title":"Certified Interventional Fidelity: Anytime-Valid, Adaptive Evaluation of Causal Claims in Mechanistic Interpretability","ref_index":16,"is_internal_anchor":true},{"citing_arxiv_id":"2606.20678","citing_title":"Beyond Importance: Interchange-Sobol Sensitivity Reveals Task-Specific Content Channels in Transformer Components","ref_index":15,"is_internal_anchor":false},{"citing_arxiv_id":"2606.12818","citing_title":"Localizing Anchoring Pathways in Language Models","ref_index":6,"is_internal_anchor":false},{"citing_arxiv_id":"2606.08044","citing_title":"When Behavioral Safety Evaluation Fails: A Representation-Level Perspective","ref_index":48,"is_internal_anchor":false},{"citing_arxiv_id":"2606.09899","citing_title":"When Attribution Patching Lies: Diagnosis and a Second-Order Correction","ref_index":7,"is_internal_anchor":false},{"citing_arxiv_id":"2605.24577","citing_title":"Polymorphism Is Rotation: Operational Mechanistic Interpretability from a Two-Layer Transformer to Pythia-70m","ref_index":7,"is_internal_anchor":false},{"citing_arxiv_id":"2605.25225","citing_title":"Transformer Field Theory: A Response-Theoretic Approach to Mechanistic Interpretability","ref_index":15,"is_internal_anchor":false},{"citing_arxiv_id":"2605.29358","citing_title":"Scaling Monosemanticity: Extracting Interpretable Features from Claude 3 Sonnet","ref_index":44,"is_internal_anchor":false},{"citing_arxiv_id":"2606.00831","citing_title":"Subliminal Learning is a LoRA Artifact","ref_index":27,"is_internal_anchor":false},{"citing_arxiv_id":"2606.06267","citing_title":"Many Circuits, One Mechanism: Input Variation and Evaluation Granularity in Circuit Discovery","ref_index":47,"is_internal_anchor":false},{"citing_arxiv_id":"2605.12770","citing_title":"WriteSAE: Sparse Autoencoders for Recurrent State","ref_index":12,"is_internal_anchor":false},{"citing_arxiv_id":"2605.12770","citing_title":"WriteSAE: Sparse Autoencoders for Recurrent State","ref_index":12,"is_internal_anchor":false},{"citing_arxiv_id":"2403.19647","citing_title":"Sparse Feature Circuits: Discovering and Editing Interpretable Causal Graphs in Language Models","ref_index":39,"is_internal_anchor":false},{"citing_arxiv_id":"2605.12207","citing_title":"Not How Many, But Which: Parameter Placement in Low-Rank Adaptation","ref_index":47,"is_internal_anchor":false},{"citing_arxiv_id":"2605.09314","citing_title":"How LLMs Are Persuaded: A Few Attention Heads, Rerouted","ref_index":32,"is_internal_anchor":false},{"citing_arxiv_id":"2605.08740","citing_title":"Causal Dimensionality of Transformer Representations: Measurement, Scaling, and Layer Structure","ref_index":13,"is_internal_anchor":false},{"citing_arxiv_id":"2604.19052","citing_title":"Cell-Based Representation of Relational Binding in Language Models","ref_index":44,"is_internal_anchor":false},{"citing_arxiv_id":"2604.08016","citing_title":"Wiring the 'Why': A Unified Taxonomy and Survey of Abductive Reasoning in LLMs","ref_index":52,"is_internal_anchor":false},{"citing_arxiv_id":"2604.13694","citing_title":"Weight Patching: Toward Source-Level Mechanistic Localization in LLMs","ref_index":24,"is_internal_anchor":false}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/ZF2G7UWTGDVOZLPQ7DYFZ5X2NJ","json":"https://pith.science/pith/ZF2G7UWTGDVOZLPQ7DYFZ5X2NJ.json","graph_json":"https://pith.science/api/pith-number/ZF2G7UWTGDVOZLPQ7DYFZ5X2NJ/graph.json","events_json":"https://pith.science/api/pith-number/ZF2G7UWTGDVOZLPQ7DYFZ5X2NJ/events.json","paper":"https://pith.science/paper/ZF2G7UWT"},"agent_actions":{"view_html":"https://pith.science/pith/ZF2G7UWTGDVOZLPQ7DYFZ5X2NJ","download_json":"https://pith.science/pith/ZF2G7UWTGDVOZLPQ7DYFZ5X2NJ.json","view_paper":"https://pith.science/paper/ZF2G7UWT","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2403.00745&json=true","fetch_graph":"https://pith.science/api/pith-number/ZF2G7UWTGDVOZLPQ7DYFZ5X2NJ/graph.json","fetch_events":"https://pith.science/api/pith-number/ZF2G7UWTGDVOZLPQ7DYFZ5X2NJ/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/ZF2G7UWTGDVOZLPQ7DYFZ5X2NJ/action/timestamp_anchor","attest_storage":"https://pith.science/pith/ZF2G7UWTGDVOZLPQ7DYFZ5X2NJ/action/storage_attestation","attest_author":"https://pith.science/pith/ZF2G7UWTGDVOZLPQ7DYFZ5X2NJ/action/author_attestation","sign_citation":"https://pith.science/pith/ZF2G7UWTGDVOZLPQ7DYFZ5X2NJ/action/citation_signature","submit_replication":"https://pith.science/pith/ZF2G7UWTGDVOZLPQ7DYFZ5X2NJ/action/replication_record"}},"created_at":"2026-07-05T07:51:08.085804+00:00","updated_at":"2026-07-05T07:51:08.085804+00:00"}