{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2025:DEORSDNWXY4N7QLYJ3CSQRXAW3","short_pith_number":"pith:DEORSDNW","schema_version":"1.0","canonical_sha256":"191d190db6be38dfc1784ec52846e0b6f66ec25ffc757b79856554a047ecce42","source":{"kind":"arxiv","id":"2509.00749","version":1},"attestation_state":"computed","paper":{"title":"Causal Interpretation of Sparse Autoencoder Features in Vision","license":"http://creativecommons.org/licenses/by-sa/4.0/","headline":"","cross_cats":["cs.AI"],"primary_cat":"cs.CV","authors_text":"Nojun Kwak, Sangyu Han, Yearim Kim","submitted_at":"2025-08-31T08:52:45Z","abstract_excerpt":"Understanding what sparse auto-encoder (SAE) features in vision transformers truly represent is usually done by inspecting the patches where a feature's activation is highest. However, self-attention mixes information across the entire image, so an activated patch often co-occurs with-but does not cause-the feature's firing. We propose Causal Feature Explanation (CaFE), which leverages Effective Receptive Field (ERF). We consider each activation of an SAE feature to be a target and apply input-attribution methods to identify the image patches that causally drive that activation. Across CLIP-Vi"},"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":"2509.00749","kind":"arxiv","version":1},"metadata":{"license":"http://creativecommons.org/licenses/by-sa/4.0/","primary_cat":"cs.CV","submitted_at":"2025-08-31T08:52:45Z","cross_cats_sorted":["cs.AI"],"title_canon_sha256":"d84032dd7993bf7adeeed88ccfaf1eabc8cf554680fb41228dc48b681f50e30d","abstract_canon_sha256":"d7950766b50592cedbf08b2d54133b03c34598bcf3ca286cc8a040956315fbbc"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T12:02:22.158170Z","signature_b64":"/mPvxwO3p52t5olkjJa6P2QHfhkD2WyROMcQtwZv35wXVM8po0/wmQt6Yvo1VmojvorV7H9dqsRuNbEGK/dcBQ==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"191d190db6be38dfc1784ec52846e0b6f66ec25ffc757b79856554a047ecce42","last_reissued_at":"2026-07-05T12:02:22.157635Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T12:02:22.157635Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Causal Interpretation of Sparse Autoencoder Features in Vision","license":"http://creativecommons.org/licenses/by-sa/4.0/","headline":"","cross_cats":["cs.AI"],"primary_cat":"cs.CV","authors_text":"Nojun Kwak, Sangyu Han, Yearim Kim","submitted_at":"2025-08-31T08:52:45Z","abstract_excerpt":"Understanding what sparse auto-encoder (SAE) features in vision transformers truly represent is usually done by inspecting the patches where a feature's activation is highest. However, self-attention mixes information across the entire image, so an activated patch often co-occurs with-but does not cause-the feature's firing. We propose Causal Feature Explanation (CaFE), which leverages Effective Receptive Field (ERF). We consider each activation of an SAE feature to be a target and apply input-attribution methods to identify the image patches that causally drive that activation. Across CLIP-Vi"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2509.00749","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/2509.00749/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":"2509.00749","created_at":"2026-07-05T12:02:22.157689+00:00"},{"alias_kind":"arxiv_version","alias_value":"2509.00749v1","created_at":"2026-07-05T12:02:22.157689+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2509.00749","created_at":"2026-07-05T12:02:22.157689+00:00"},{"alias_kind":"pith_short_12","alias_value":"DEORSDNWXY4N","created_at":"2026-07-05T12:02:22.157689+00:00"},{"alias_kind":"pith_short_16","alias_value":"DEORSDNWXY4N7QLY","created_at":"2026-07-05T12:02:22.157689+00:00"},{"alias_kind":"pith_short_8","alias_value":"DEORSDNW","created_at":"2026-07-05T12:02:22.157689+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":2,"internal_anchor_count":0,"sample":[{"citing_arxiv_id":"2605.15961","citing_title":"Sparse Autoencoders enable Robust and Interpretable Fine-tuning of CLIP models","ref_index":10,"is_internal_anchor":false},{"citing_arxiv_id":"2604.05724","citing_title":"Beyond Semantics: Disentangling Information Scope in Sparse Autoencoders for CLIP","ref_index":17,"is_internal_anchor":false}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/DEORSDNWXY4N7QLYJ3CSQRXAW3","json":"https://pith.science/pith/DEORSDNWXY4N7QLYJ3CSQRXAW3.json","graph_json":"https://pith.science/api/pith-number/DEORSDNWXY4N7QLYJ3CSQRXAW3/graph.json","events_json":"https://pith.science/api/pith-number/DEORSDNWXY4N7QLYJ3CSQRXAW3/events.json","paper":"https://pith.science/paper/DEORSDNW"},"agent_actions":{"view_html":"https://pith.science/pith/DEORSDNWXY4N7QLYJ3CSQRXAW3","download_json":"https://pith.science/pith/DEORSDNWXY4N7QLYJ3CSQRXAW3.json","view_paper":"https://pith.science/paper/DEORSDNW","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2509.00749&json=true","fetch_graph":"https://pith.science/api/pith-number/DEORSDNWXY4N7QLYJ3CSQRXAW3/graph.json","fetch_events":"https://pith.science/api/pith-number/DEORSDNWXY4N7QLYJ3CSQRXAW3/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/DEORSDNWXY4N7QLYJ3CSQRXAW3/action/timestamp_anchor","attest_storage":"https://pith.science/pith/DEORSDNWXY4N7QLYJ3CSQRXAW3/action/storage_attestation","attest_author":"https://pith.science/pith/DEORSDNWXY4N7QLYJ3CSQRXAW3/action/author_attestation","sign_citation":"https://pith.science/pith/DEORSDNWXY4N7QLYJ3CSQRXAW3/action/citation_signature","submit_replication":"https://pith.science/pith/DEORSDNWXY4N7QLYJ3CSQRXAW3/action/replication_record"}},"created_at":"2026-07-05T12:02:22.157689+00:00","updated_at":"2026-07-05T12:02:22.157689+00:00"}