{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2025:XV7NRO3OM3FW22QTTB3H26HT34","short_pith_number":"pith:XV7NRO3O","schema_version":"1.0","canonical_sha256":"bd7ed8bb6e66cb6d6a1398767d78f3df1d6d1e102e91a7077fb83ff03fda8e55","source":{"kind":"arxiv","id":"2501.18823","version":2},"attestation_state":"computed","paper":{"title":"Transcoders Beat Sparse Autoencoders for Interpretability","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":[],"primary_cat":"cs.LG","authors_text":"Gon\\c{c}alo Paulo, Nora Belrose, Stepan Shabalin","submitted_at":"2025-01-31T00:36:30Z","abstract_excerpt":"Sparse autoencoders (SAEs) extract human-interpretable features from deep neural networks by transforming their activations into a sparse, higher dimensional latent space, and then reconstructing the activations from these latents. Transcoders are similar to SAEs, but they are trained to reconstruct the output of a component of a deep network given its input. In this work, we compare the features found by transcoders and SAEs trained on the same model and data, finding that transcoder features are significantly more interpretable. We also propose skip transcoders, which add an affine skip conn"},"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":"2501.18823","kind":"arxiv","version":2},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.LG","submitted_at":"2025-01-31T00:36:30Z","cross_cats_sorted":[],"title_canon_sha256":"7e7840ef68f5bb95c53fcbda2a6f2d221ccfda4f07ad068bf6f50d147320bd0e","abstract_canon_sha256":"04668569de8556d51bd7dff60fca22bc484084079fcb7e82037b12674d97d5db"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T10:13:07.976627Z","signature_b64":"ukEHGaH/2lws4T/97hGuDVWGk4Ye+IUFAFsSJCqHcNbocqeq2uYjacFElwrISVn7BvBZNA76yqlNNL3xUsh4Bg==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"bd7ed8bb6e66cb6d6a1398767d78f3df1d6d1e102e91a7077fb83ff03fda8e55","last_reissued_at":"2026-07-05T10:13:07.976167Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T10:13:07.976167Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Transcoders Beat Sparse Autoencoders for Interpretability","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":[],"primary_cat":"cs.LG","authors_text":"Gon\\c{c}alo Paulo, Nora Belrose, Stepan Shabalin","submitted_at":"2025-01-31T00:36:30Z","abstract_excerpt":"Sparse autoencoders (SAEs) extract human-interpretable features from deep neural networks by transforming their activations into a sparse, higher dimensional latent space, and then reconstructing the activations from these latents. Transcoders are similar to SAEs, but they are trained to reconstruct the output of a component of a deep network given its input. In this work, we compare the features found by transcoders and SAEs trained on the same model and data, finding that transcoder features are significantly more interpretable. We also propose skip transcoders, which add an affine skip conn"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2501.18823","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/2501.18823/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":"2501.18823","created_at":"2026-07-05T10:13:07.976222+00:00"},{"alias_kind":"arxiv_version","alias_value":"2501.18823v2","created_at":"2026-07-05T10:13:07.976222+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2501.18823","created_at":"2026-07-05T10:13:07.976222+00:00"},{"alias_kind":"pith_short_12","alias_value":"XV7NRO3OM3FW","created_at":"2026-07-05T10:13:07.976222+00:00"},{"alias_kind":"pith_short_16","alias_value":"XV7NRO3OM3FW22QT","created_at":"2026-07-05T10:13:07.976222+00:00"},{"alias_kind":"pith_short_8","alias_value":"XV7NRO3O","created_at":"2026-07-05T10:13:07.976222+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":8,"internal_anchor_count":0,"sample":[{"citing_arxiv_id":"2606.09940","citing_title":"Interactions Between Crosscoder Features: A Compact Proofs Perspective","ref_index":31,"is_internal_anchor":false},{"citing_arxiv_id":"2605.21849","citing_title":"Geometry-Adaptive Explainer for Faithful Dictionary-Based Interpretability under Distribution Shift","ref_index":22,"is_internal_anchor":false},{"citing_arxiv_id":"2605.12770","citing_title":"WriteSAE: Sparse Autoencoders for Recurrent State","ref_index":35,"is_internal_anchor":false},{"citing_arxiv_id":"2605.12770","citing_title":"WriteSAE: Sparse Autoencoders for Recurrent State","ref_index":82,"is_internal_anchor":false},{"citing_arxiv_id":"2511.01680","citing_title":"Making Interpretable Discoveries from Unstructured Data: A High-Dimensional Multiple Hypothesis Testing Approach","ref_index":38,"is_internal_anchor":false},{"citing_arxiv_id":"2605.12770","citing_title":"WriteSAE: Sparse Autoencoders for Recurrent State","ref_index":82,"is_internal_anchor":false},{"citing_arxiv_id":"2605.12770","citing_title":"WriteSAE: Sparse Autoencoders for Recurrent State","ref_index":82,"is_internal_anchor":false},{"citing_arxiv_id":"2604.06495","citing_title":"Improving Robustness In Sparse Autoencoders via Masked Regularization","ref_index":12,"is_internal_anchor":false}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/XV7NRO3OM3FW22QTTB3H26HT34","json":"https://pith.science/pith/XV7NRO3OM3FW22QTTB3H26HT34.json","graph_json":"https://pith.science/api/pith-number/XV7NRO3OM3FW22QTTB3H26HT34/graph.json","events_json":"https://pith.science/api/pith-number/XV7NRO3OM3FW22QTTB3H26HT34/events.json","paper":"https://pith.science/paper/XV7NRO3O"},"agent_actions":{"view_html":"https://pith.science/pith/XV7NRO3OM3FW22QTTB3H26HT34","download_json":"https://pith.science/pith/XV7NRO3OM3FW22QTTB3H26HT34.json","view_paper":"https://pith.science/paper/XV7NRO3O","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2501.18823&json=true","fetch_graph":"https://pith.science/api/pith-number/XV7NRO3OM3FW22QTTB3H26HT34/graph.json","fetch_events":"https://pith.science/api/pith-number/XV7NRO3OM3FW22QTTB3H26HT34/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/XV7NRO3OM3FW22QTTB3H26HT34/action/timestamp_anchor","attest_storage":"https://pith.science/pith/XV7NRO3OM3FW22QTTB3H26HT34/action/storage_attestation","attest_author":"https://pith.science/pith/XV7NRO3OM3FW22QTTB3H26HT34/action/author_attestation","sign_citation":"https://pith.science/pith/XV7NRO3OM3FW22QTTB3H26HT34/action/citation_signature","submit_replication":"https://pith.science/pith/XV7NRO3OM3FW22QTTB3H26HT34/action/replication_record"}},"created_at":"2026-07-05T10:13:07.976222+00:00","updated_at":"2026-07-05T10:13:07.976222+00:00"}