{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2024:HPG2X7MDOHOPILIZYYV2TQUOYQ","short_pith_number":"pith:HPG2X7MD","schema_version":"1.0","canonical_sha256":"3bcdabfd8371dcf42d19c62ba9c28ec40ff610eeb5b4e30622f4da337f59c170","source":{"kind":"arxiv","id":"2409.04478","version":1},"attestation_state":"computed","paper":{"title":"Evaluating Open-Source Sparse Autoencoders on Disentangling Factual Knowledge in GPT-2 Small","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.AI","cs.NE"],"primary_cat":"cs.LG","authors_text":"Atticus Geiger, Maheep Chaudhary","submitted_at":"2024-09-05T18:00:37Z","abstract_excerpt":"A popular new method in mechanistic interpretability is to train high-dimensional sparse autoencoders (SAEs) on neuron activations and use SAE features as the atomic units of analysis. However, the body of evidence on whether SAE feature spaces are useful for causal analysis is underdeveloped. In this work, we use the RAVEL benchmark to evaluate whether SAEs trained on hidden representations of GPT-2 small have sets of features that separately mediate knowledge of which country a city is in and which continent it is in. We evaluate four open-source SAEs for GPT-2 small against each other, with"},"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":"2409.04478","kind":"arxiv","version":1},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.LG","submitted_at":"2024-09-05T18:00:37Z","cross_cats_sorted":["cs.AI","cs.NE"],"title_canon_sha256":"4e1ea02ec6a0b8c2dfdfabddca1cb05a7341b2d36a09e698a8c9a7f3eb55ae8c","abstract_canon_sha256":"d5aab075b7923f04b43f00ab6befe45e7620b24c9d2ff9f38affbcd4306fb8de"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T09:04:12.676501Z","signature_b64":"4M38dxzCkvjGoaaD5f70wYbxpHsBarsG6SCBmzRSvPAoLYIKHdN41u8OzfSWwM4s2wGAU295x0tFV8taQZHkBw==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"3bcdabfd8371dcf42d19c62ba9c28ec40ff610eeb5b4e30622f4da337f59c170","last_reissued_at":"2026-07-05T09:04:12.676080Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T09:04:12.676080Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Evaluating Open-Source Sparse Autoencoders on Disentangling Factual Knowledge in GPT-2 Small","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.AI","cs.NE"],"primary_cat":"cs.LG","authors_text":"Atticus Geiger, Maheep Chaudhary","submitted_at":"2024-09-05T18:00:37Z","abstract_excerpt":"A popular new method in mechanistic interpretability is to train high-dimensional sparse autoencoders (SAEs) on neuron activations and use SAE features as the atomic units of analysis. However, the body of evidence on whether SAE feature spaces are useful for causal analysis is underdeveloped. In this work, we use the RAVEL benchmark to evaluate whether SAEs trained on hidden representations of GPT-2 small have sets of features that separately mediate knowledge of which country a city is in and which continent it is in. We evaluate four open-source SAEs for GPT-2 small against each other, with"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2409.04478","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/2409.04478/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":"2409.04478","created_at":"2026-07-05T09:04:12.676139+00:00"},{"alias_kind":"arxiv_version","alias_value":"2409.04478v1","created_at":"2026-07-05T09:04:12.676139+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2409.04478","created_at":"2026-07-05T09:04:12.676139+00:00"},{"alias_kind":"pith_short_12","alias_value":"HPG2X7MDOHOP","created_at":"2026-07-05T09:04:12.676139+00:00"},{"alias_kind":"pith_short_16","alias_value":"HPG2X7MDOHOPILIZ","created_at":"2026-07-05T09:04:12.676139+00:00"},{"alias_kind":"pith_short_8","alias_value":"HPG2X7MD","created_at":"2026-07-05T09:04:12.676139+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":5,"internal_anchor_count":0,"sample":[{"citing_arxiv_id":"2606.18897","citing_title":"SAERec: Constructing Fine-grained Interpretable Intents Priors via Sparse Autoencoders for Recommendation","ref_index":6,"is_internal_anchor":false},{"citing_arxiv_id":"2606.07720","citing_title":"Why Limit the Residual Stream to Layers and Not Tokens? Persistent Memory for Continuous Latent Reasoning","ref_index":35,"is_internal_anchor":false},{"citing_arxiv_id":"2605.28860","citing_title":"Mechanistic origins of catastrophic forgetting: why RL preserves circuits better than SFT?","ref_index":3,"is_internal_anchor":false},{"citing_arxiv_id":"2605.06979","citing_title":"PLOT: Progressive Localization via Optimal Transport in Neural Causal Abstraction","ref_index":2,"is_internal_anchor":false},{"citing_arxiv_id":"2605.06979","citing_title":"PLOT: Progressive Localization via Optimal Transport in Neural Causal Abstraction","ref_index":3,"is_internal_anchor":false}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/HPG2X7MDOHOPILIZYYV2TQUOYQ","json":"https://pith.science/pith/HPG2X7MDOHOPILIZYYV2TQUOYQ.json","graph_json":"https://pith.science/api/pith-number/HPG2X7MDOHOPILIZYYV2TQUOYQ/graph.json","events_json":"https://pith.science/api/pith-number/HPG2X7MDOHOPILIZYYV2TQUOYQ/events.json","paper":"https://pith.science/paper/HPG2X7MD"},"agent_actions":{"view_html":"https://pith.science/pith/HPG2X7MDOHOPILIZYYV2TQUOYQ","download_json":"https://pith.science/pith/HPG2X7MDOHOPILIZYYV2TQUOYQ.json","view_paper":"https://pith.science/paper/HPG2X7MD","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2409.04478&json=true","fetch_graph":"https://pith.science/api/pith-number/HPG2X7MDOHOPILIZYYV2TQUOYQ/graph.json","fetch_events":"https://pith.science/api/pith-number/HPG2X7MDOHOPILIZYYV2TQUOYQ/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/HPG2X7MDOHOPILIZYYV2TQUOYQ/action/timestamp_anchor","attest_storage":"https://pith.science/pith/HPG2X7MDOHOPILIZYYV2TQUOYQ/action/storage_attestation","attest_author":"https://pith.science/pith/HPG2X7MDOHOPILIZYYV2TQUOYQ/action/author_attestation","sign_citation":"https://pith.science/pith/HPG2X7MDOHOPILIZYYV2TQUOYQ/action/citation_signature","submit_replication":"https://pith.science/pith/HPG2X7MDOHOPILIZYYV2TQUOYQ/action/replication_record"}},"created_at":"2026-07-05T09:04:12.676139+00:00","updated_at":"2026-07-05T09:04:12.676139+00:00"}