{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2026:LS2OBQ5JNBW4PPPI4HOB4AL4VK","short_pith_number":"pith:LS2OBQ5J","schema_version":"1.0","canonical_sha256":"5cb4e0c3a9686dc7bde8e1dc1e017caabc38bb75d8f7a2a94a8099f629316cf3","source":{"kind":"arxiv","id":"2606.06333","version":1},"attestation_state":"computed","paper":{"title":"Subspace-Aware Sparse Autoencoders for Effective Mechanistic Interpretability","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.AI"],"primary_cat":"cs.LG","authors_text":"Mehrdad Mahdavi, Seyed Arshan Dalili","submitted_at":"2026-06-04T16:08:25Z","abstract_excerpt":"Sparse Autoencoders (SAEs) are widely used for mechanistic interpretability in large language models, yet their formulation assigns each latent feature a single decoder direction, implicitly assuming features to be one-dimensional. We show that this assumption mismatches with the multi-dimensional structure of model features, provably inducing feature splitting through two distinct mechanisms. Geometrically, reconstructing a feature of intrinsic dimension $d_i \\ge 2$ to error $\\varepsilon$ with single-direction decoders forces a number of atoms that is exponential in $d_i$. From an end-to-end "},"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":"2606.06333","kind":"arxiv","version":1},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.LG","submitted_at":"2026-06-04T16:08:25Z","cross_cats_sorted":["cs.AI"],"title_canon_sha256":"55737eb27ce4c0e4760095286d443a2cea37655a02e85837faa04479c0849a42","abstract_canon_sha256":"08b9af34127efdc07cb50ee44f0ae1b0c3b417f7b2256844a21dad4a1a80f3be"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-06-05T01:15:42.106371Z","signature_b64":"5mep71X9ij/WUCwIZM/EuwCxCjvGnEmjiFP2Yso2TCqVKmB0fEPDteRNd/dOYxkUuPOl5oygX39urNlIohkBCA==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"5cb4e0c3a9686dc7bde8e1dc1e017caabc38bb75d8f7a2a94a8099f629316cf3","last_reissued_at":"2026-06-05T01:15:42.105956Z","signature_status":"signed_v1","first_computed_at":"2026-06-05T01:15:42.105956Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Subspace-Aware Sparse Autoencoders for Effective Mechanistic Interpretability","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.AI"],"primary_cat":"cs.LG","authors_text":"Mehrdad Mahdavi, Seyed Arshan Dalili","submitted_at":"2026-06-04T16:08:25Z","abstract_excerpt":"Sparse Autoencoders (SAEs) are widely used for mechanistic interpretability in large language models, yet their formulation assigns each latent feature a single decoder direction, implicitly assuming features to be one-dimensional. We show that this assumption mismatches with the multi-dimensional structure of model features, provably inducing feature splitting through two distinct mechanisms. Geometrically, reconstructing a feature of intrinsic dimension $d_i \\ge 2$ to error $\\varepsilon$ with single-direction decoders forces a number of atoms that is exponential in $d_i$. From an end-to-end "},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2606.06333","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/2606.06333/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":"2606.06333","created_at":"2026-06-05T01:15:42.106022+00:00"},{"alias_kind":"arxiv_version","alias_value":"2606.06333v1","created_at":"2026-06-05T01:15:42.106022+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2606.06333","created_at":"2026-06-05T01:15:42.106022+00:00"},{"alias_kind":"pith_short_12","alias_value":"LS2OBQ5JNBW4","created_at":"2026-06-05T01:15:42.106022+00:00"},{"alias_kind":"pith_short_16","alias_value":"LS2OBQ5JNBW4PPPI","created_at":"2026-06-05T01:15:42.106022+00:00"},{"alias_kind":"pith_short_8","alias_value":"LS2OBQ5J","created_at":"2026-06-05T01:15:42.106022+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":0,"internal_anchor_count":0,"sample":[]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/LS2OBQ5JNBW4PPPI4HOB4AL4VK","json":"https://pith.science/pith/LS2OBQ5JNBW4PPPI4HOB4AL4VK.json","graph_json":"https://pith.science/api/pith-number/LS2OBQ5JNBW4PPPI4HOB4AL4VK/graph.json","events_json":"https://pith.science/api/pith-number/LS2OBQ5JNBW4PPPI4HOB4AL4VK/events.json","paper":"https://pith.science/paper/LS2OBQ5J"},"agent_actions":{"view_html":"https://pith.science/pith/LS2OBQ5JNBW4PPPI4HOB4AL4VK","download_json":"https://pith.science/pith/LS2OBQ5JNBW4PPPI4HOB4AL4VK.json","view_paper":"https://pith.science/paper/LS2OBQ5J","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2606.06333&json=true","fetch_graph":"https://pith.science/api/pith-number/LS2OBQ5JNBW4PPPI4HOB4AL4VK/graph.json","fetch_events":"https://pith.science/api/pith-number/LS2OBQ5JNBW4PPPI4HOB4AL4VK/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/LS2OBQ5JNBW4PPPI4HOB4AL4VK/action/timestamp_anchor","attest_storage":"https://pith.science/pith/LS2OBQ5JNBW4PPPI4HOB4AL4VK/action/storage_attestation","attest_author":"https://pith.science/pith/LS2OBQ5JNBW4PPPI4HOB4AL4VK/action/author_attestation","sign_citation":"https://pith.science/pith/LS2OBQ5JNBW4PPPI4HOB4AL4VK/action/citation_signature","submit_replication":"https://pith.science/pith/LS2OBQ5JNBW4PPPI4HOB4AL4VK/action/replication_record"}},"created_at":"2026-06-05T01:15:42.106022+00:00","updated_at":"2026-06-05T01:15:42.106022+00:00"}