{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2024:6INO655OCX65GAVNQTOXMX22MT","short_pith_number":"pith:6INO655O","schema_version":"1.0","canonical_sha256":"f21aef77ae15fdd302ad84dd765f5a64c818e43398ee629d50b9928c6ae8e75a","source":{"kind":"arxiv","id":"2411.18895","version":1},"attestation_state":"computed","paper":{"title":"Evaluating Sparse Autoencoders on Targeted Concept Erasure Tasks","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.CL"],"primary_cat":"cs.LG","authors_text":"Adam Karvonen, Can Rager, Neel Nanda, Samuel Marks","submitted_at":"2024-11-28T03:58:48Z","abstract_excerpt":"Sparse Autoencoders (SAEs) are an interpretability technique aimed at decomposing neural network activations into interpretable units. However, a major bottleneck for SAE development has been the lack of high-quality performance metrics, with prior work largely relying on unsupervised proxies. In this work, we introduce a family of evaluations based on SHIFT, a downstream task from Marks et al. (Sparse Feature Circuits, 2024) in which spurious cues are removed from a classifier by ablating SAE features judged to be task-irrelevant by a human annotator. We adapt SHIFT into an automated metric o"},"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":"2411.18895","kind":"arxiv","version":1},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.LG","submitted_at":"2024-11-28T03:58:48Z","cross_cats_sorted":["cs.CL"],"title_canon_sha256":"658586de7bf322a9dacf226dc73d8801d749b72572c851092d5134dc4ec5120c","abstract_canon_sha256":"4326b5137cd0de75e27dd7b785893b6ace4d143805025146feb7aab552373e7a"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T09:41:53.364158Z","signature_b64":"15B8g3aYKeQje7Y5UDRqO62FBTRiFD5nu7xqOPiAbMszqPpxXzZUzhhFL87xKBT5SXYviaZbTBcHANEaHBNUBA==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"f21aef77ae15fdd302ad84dd765f5a64c818e43398ee629d50b9928c6ae8e75a","last_reissued_at":"2026-07-05T09:41:53.363675Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T09:41:53.363675Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Evaluating Sparse Autoencoders on Targeted Concept Erasure Tasks","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.CL"],"primary_cat":"cs.LG","authors_text":"Adam Karvonen, Can Rager, Neel Nanda, Samuel Marks","submitted_at":"2024-11-28T03:58:48Z","abstract_excerpt":"Sparse Autoencoders (SAEs) are an interpretability technique aimed at decomposing neural network activations into interpretable units. However, a major bottleneck for SAE development has been the lack of high-quality performance metrics, with prior work largely relying on unsupervised proxies. In this work, we introduce a family of evaluations based on SHIFT, a downstream task from Marks et al. (Sparse Feature Circuits, 2024) in which spurious cues are removed from a classifier by ablating SAE features judged to be task-irrelevant by a human annotator. We adapt SHIFT into an automated metric o"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2411.18895","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/2411.18895/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":"2411.18895","created_at":"2026-07-05T09:41:53.363739+00:00"},{"alias_kind":"arxiv_version","alias_value":"2411.18895v1","created_at":"2026-07-05T09:41:53.363739+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2411.18895","created_at":"2026-07-05T09:41:53.363739+00:00"},{"alias_kind":"pith_short_12","alias_value":"6INO655OCX65","created_at":"2026-07-05T09:41:53.363739+00:00"},{"alias_kind":"pith_short_16","alias_value":"6INO655OCX65GAVN","created_at":"2026-07-05T09:41:53.363739+00:00"},{"alias_kind":"pith_short_8","alias_value":"6INO655O","created_at":"2026-07-05T09:41:53.363739+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":3,"internal_anchor_count":0,"sample":[{"citing_arxiv_id":"2605.18629","citing_title":"Aligned Training: A Parameter-Free Method to Improve Feature Quality and Stability of Sparse Autoencoders (SAE)","ref_index":19,"is_internal_anchor":false},{"citing_arxiv_id":"2605.27819","citing_title":"ReSAE: Residualized Sparse Autoencoders for Multi-Layer Transformer Interventions","ref_index":4,"is_internal_anchor":false},{"citing_arxiv_id":"2605.18629","citing_title":"Aligned Training: A Parameter-Free Method to Improve Feature Quality and Stability of Sparse Autoencoders (SAE)","ref_index":17,"is_internal_anchor":false}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/6INO655OCX65GAVNQTOXMX22MT","json":"https://pith.science/pith/6INO655OCX65GAVNQTOXMX22MT.json","graph_json":"https://pith.science/api/pith-number/6INO655OCX65GAVNQTOXMX22MT/graph.json","events_json":"https://pith.science/api/pith-number/6INO655OCX65GAVNQTOXMX22MT/events.json","paper":"https://pith.science/paper/6INO655O"},"agent_actions":{"view_html":"https://pith.science/pith/6INO655OCX65GAVNQTOXMX22MT","download_json":"https://pith.science/pith/6INO655OCX65GAVNQTOXMX22MT.json","view_paper":"https://pith.science/paper/6INO655O","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2411.18895&json=true","fetch_graph":"https://pith.science/api/pith-number/6INO655OCX65GAVNQTOXMX22MT/graph.json","fetch_events":"https://pith.science/api/pith-number/6INO655OCX65GAVNQTOXMX22MT/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/6INO655OCX65GAVNQTOXMX22MT/action/timestamp_anchor","attest_storage":"https://pith.science/pith/6INO655OCX65GAVNQTOXMX22MT/action/storage_attestation","attest_author":"https://pith.science/pith/6INO655OCX65GAVNQTOXMX22MT/action/author_attestation","sign_citation":"https://pith.science/pith/6INO655OCX65GAVNQTOXMX22MT/action/citation_signature","submit_replication":"https://pith.science/pith/6INO655OCX65GAVNQTOXMX22MT/action/replication_record"}},"created_at":"2026-07-05T09:41:53.363739+00:00","updated_at":"2026-07-05T09:41:53.363739+00:00"}