{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2025:MYMZXOWGTDLSAKGS7KC4KQMDAF","short_pith_number":"pith:MYMZXOWG","schema_version":"1.0","canonical_sha256":"66199bbac698d72028d2fa85c541830153b3c791b51dcc3bf3d8309ecf506450","source":{"kind":"arxiv","id":"2502.17332","version":1},"attestation_state":"computed","paper":{"title":"Tokenized SAEs: Disentangling SAE Reconstructions","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":[],"primary_cat":"cs.LG","authors_text":"Daniel Wilhelm, Thomas Dooms","submitted_at":"2025-02-24T17:04:24Z","abstract_excerpt":"Sparse auto-encoders (SAEs) have become a prevalent tool for interpreting language models' inner workings. However, it is unknown how tightly SAE features correspond to computationally important directions in the model. This work empirically shows that many RES-JB SAE features predominantly correspond to simple input statistics. We hypothesize this is caused by a large class imbalance in training data combined with a lack of complex error signals. To reduce this behavior, we propose a method that disentangles token reconstruction from feature reconstruction. This improvement is achieved by int"},"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":"2502.17332","kind":"arxiv","version":1},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.LG","submitted_at":"2025-02-24T17:04:24Z","cross_cats_sorted":[],"title_canon_sha256":"870d07414773ba589d1a1381e52931ed8234deacb096a6a159213d78fe6d9fef","abstract_canon_sha256":"3dff60af18bca4e88de940d1b6fd953c1775b4887a56759458972a77c32e7ce3"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T10:19:08.989547Z","signature_b64":"D8lPn4W8YWN1zG2bNxRCjZ3yQrVTiDZAu9xkz5eGQK27Pu5MiR1T3c9SoEO6iW1iJb1E+OsAXsIFrVRtTH+qAg==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"66199bbac698d72028d2fa85c541830153b3c791b51dcc3bf3d8309ecf506450","last_reissued_at":"2026-07-05T10:19:08.989090Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T10:19:08.989090Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Tokenized SAEs: Disentangling SAE Reconstructions","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":[],"primary_cat":"cs.LG","authors_text":"Daniel Wilhelm, Thomas Dooms","submitted_at":"2025-02-24T17:04:24Z","abstract_excerpt":"Sparse auto-encoders (SAEs) have become a prevalent tool for interpreting language models' inner workings. However, it is unknown how tightly SAE features correspond to computationally important directions in the model. This work empirically shows that many RES-JB SAE features predominantly correspond to simple input statistics. We hypothesize this is caused by a large class imbalance in training data combined with a lack of complex error signals. To reduce this behavior, we propose a method that disentangles token reconstruction from feature reconstruction. This improvement is achieved by int"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2502.17332","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/2502.17332/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":"2502.17332","created_at":"2026-07-05T10:19:08.989144+00:00"},{"alias_kind":"arxiv_version","alias_value":"2502.17332v1","created_at":"2026-07-05T10:19:08.989144+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2502.17332","created_at":"2026-07-05T10:19:08.989144+00:00"},{"alias_kind":"pith_short_12","alias_value":"MYMZXOWGTDLS","created_at":"2026-07-05T10:19:08.989144+00:00"},{"alias_kind":"pith_short_16","alias_value":"MYMZXOWGTDLSAKGS","created_at":"2026-07-05T10:19:08.989144+00:00"},{"alias_kind":"pith_short_8","alias_value":"MYMZXOWG","created_at":"2026-07-05T10:19:08.989144+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":2,"internal_anchor_count":1,"sample":[{"citing_arxiv_id":"2607.06114","citing_title":"x-Prediction Is All You Need:Training-Free Accelerated Generation via Endpoint Decodability","ref_index":28,"is_internal_anchor":true},{"citing_arxiv_id":"2605.06124","citing_title":"P-Guide: Parameter-Efficient Prior Steering for Single-Pass CFG Inference","ref_index":45,"is_internal_anchor":false}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/MYMZXOWGTDLSAKGS7KC4KQMDAF","json":"https://pith.science/pith/MYMZXOWGTDLSAKGS7KC4KQMDAF.json","graph_json":"https://pith.science/api/pith-number/MYMZXOWGTDLSAKGS7KC4KQMDAF/graph.json","events_json":"https://pith.science/api/pith-number/MYMZXOWGTDLSAKGS7KC4KQMDAF/events.json","paper":"https://pith.science/paper/MYMZXOWG"},"agent_actions":{"view_html":"https://pith.science/pith/MYMZXOWGTDLSAKGS7KC4KQMDAF","download_json":"https://pith.science/pith/MYMZXOWGTDLSAKGS7KC4KQMDAF.json","view_paper":"https://pith.science/paper/MYMZXOWG","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2502.17332&json=true","fetch_graph":"https://pith.science/api/pith-number/MYMZXOWGTDLSAKGS7KC4KQMDAF/graph.json","fetch_events":"https://pith.science/api/pith-number/MYMZXOWGTDLSAKGS7KC4KQMDAF/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/MYMZXOWGTDLSAKGS7KC4KQMDAF/action/timestamp_anchor","attest_storage":"https://pith.science/pith/MYMZXOWGTDLSAKGS7KC4KQMDAF/action/storage_attestation","attest_author":"https://pith.science/pith/MYMZXOWGTDLSAKGS7KC4KQMDAF/action/author_attestation","sign_citation":"https://pith.science/pith/MYMZXOWGTDLSAKGS7KC4KQMDAF/action/citation_signature","submit_replication":"https://pith.science/pith/MYMZXOWGTDLSAKGS7KC4KQMDAF/action/replication_record"}},"created_at":"2026-07-05T10:19:08.989144+00:00","updated_at":"2026-07-05T10:19:08.989144+00:00"}