{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2025:PWLSA3CRXYWS3I6LKJDKY7HZYL","short_pith_number":"pith:PWLSA3CR","schema_version":"1.0","canonical_sha256":"7d97206c51be2d2da3cb5246ac7cf9c2c30ad9daca29cce414bba745e44c5294","source":{"kind":"arxiv","id":"2502.04878","version":1},"attestation_state":"computed","paper":{"title":"Sparse Autoencoders Do Not Find Canonical Units of Analysis","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.AI"],"primary_cat":"cs.LG","authors_text":"Bart Bussmann, Curt Tigges, Joseph Bloom, Lee Sharkey, Michael Pearce, Neel Nanda, Noura Al Moubayed, Patrick Leask","submitted_at":"2025-02-07T12:33:08Z","abstract_excerpt":"A common goal of mechanistic interpretability is to decompose the activations of neural networks into features: interpretable properties of the input computed by the model. Sparse autoencoders (SAEs) are a popular method for finding these features in LLMs, and it has been postulated that they can be used to find a \\textit{canonical} set of units: a unique and complete list of atomic features. We cast doubt on this belief using two novel techniques: SAE stitching to show they are incomplete, and meta-SAEs to show they are not atomic. SAE stitching involves inserting or swapping latents from a l"},"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":"2502.04878","kind":"arxiv","version":1},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.LG","submitted_at":"2025-02-07T12:33:08Z","cross_cats_sorted":["cs.AI"],"title_canon_sha256":"3c106e5b88405707b6604c676d4120f602c2c6197a9a137c573d79f576d139f4","abstract_canon_sha256":"5408f3d05f176c228d4b5901896fd743ced755050cbc553b66ea40b45aa6510d"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T10:10:58.955565Z","signature_b64":"OGnEjNxFfqGKJ3YpWtk7Lm75L3eBt7Wm2hKQwwggeXjulRkMKedGIS6kzR+ulSYcs/lvP4kX5c5wIXYqfzE3CA==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"7d97206c51be2d2da3cb5246ac7cf9c2c30ad9daca29cce414bba745e44c5294","last_reissued_at":"2026-07-05T10:10:58.954987Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T10:10:58.954987Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Sparse Autoencoders Do Not Find Canonical Units of Analysis","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.AI"],"primary_cat":"cs.LG","authors_text":"Bart Bussmann, Curt Tigges, Joseph Bloom, Lee Sharkey, Michael Pearce, Neel Nanda, Noura Al Moubayed, Patrick Leask","submitted_at":"2025-02-07T12:33:08Z","abstract_excerpt":"A common goal of mechanistic interpretability is to decompose the activations of neural networks into features: interpretable properties of the input computed by the model. Sparse autoencoders (SAEs) are a popular method for finding these features in LLMs, and it has been postulated that they can be used to find a \\textit{canonical} set of units: a unique and complete list of atomic features. We cast doubt on this belief using two novel techniques: SAE stitching to show they are incomplete, and meta-SAEs to show they are not atomic. SAE stitching involves inserting or swapping latents from a l"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2502.04878","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/2502.04878/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":"2502.04878","created_at":"2026-07-05T10:10:58.955049+00:00"},{"alias_kind":"arxiv_version","alias_value":"2502.04878v1","created_at":"2026-07-05T10:10:58.955049+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2502.04878","created_at":"2026-07-05T10:10:58.955049+00:00"},{"alias_kind":"pith_short_12","alias_value":"PWLSA3CRXYWS","created_at":"2026-07-05T10:10:58.955049+00:00"},{"alias_kind":"pith_short_16","alias_value":"PWLSA3CRXYWS3I6L","created_at":"2026-07-05T10:10:58.955049+00:00"},{"alias_kind":"pith_short_8","alias_value":"PWLSA3CR","created_at":"2026-07-05T10:10:58.955049+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":15,"internal_anchor_count":0,"sample":[{"citing_arxiv_id":"2606.22994","citing_title":"Do Sparse Autoencoders Learn Meaningful Concept Hierarchies?","ref_index":24,"is_internal_anchor":false},{"citing_arxiv_id":"2606.20347","citing_title":"Critical Percolation as a Synthetic Data Model for Interpretability","ref_index":30,"is_internal_anchor":false},{"citing_arxiv_id":"2606.32008","citing_title":"Surrogate Fidelity: When Can Open LLMs Explain Closed Ones?","ref_index":41,"is_internal_anchor":false},{"citing_arxiv_id":"2606.15054","citing_title":"Size Doesn't Matter: Cosine-Scored Sparse Autoencoders","ref_index":27,"is_internal_anchor":false},{"citing_arxiv_id":"2605.28149","citing_title":"Sign-Aware Gated Sparse Autoencoders: Modeling Anticorrelated Features with Bi-Jump-ReLU Activations","ref_index":22,"is_internal_anchor":false},{"citing_arxiv_id":"2606.26396","citing_title":"At the Edge of Understanding: Sparse Autoencoders Trace The Limits of Transformer Generalization","ref_index":48,"is_internal_anchor":false},{"citing_arxiv_id":"2605.12770","citing_title":"WriteSAE: Sparse Autoencoders for Recurrent State","ref_index":94,"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":36,"is_internal_anchor":false},{"citing_arxiv_id":"2605.12770","citing_title":"WriteSAE: Sparse Autoencoders for Recurrent State","ref_index":94,"is_internal_anchor":false},{"citing_arxiv_id":"2605.12770","citing_title":"WriteSAE: Sparse Autoencoders for Recurrent State","ref_index":94,"is_internal_anchor":false},{"citing_arxiv_id":"2605.12874","citing_title":"Descriptive Collision in Sparse Autoencoder Auto-Interpretability: When One Explanation Describes Many Features","ref_index":16,"is_internal_anchor":false},{"citing_arxiv_id":"2605.12122","citing_title":"Disentangled Sparse Representations for Concept-Separated Diffusion Unlearning","ref_index":24,"is_internal_anchor":false},{"citing_arxiv_id":"2605.07922","citing_title":"Tree SAE: Learning Hierarchical Feature Structures in Sparse Autoencoders","ref_index":13,"is_internal_anchor":false},{"citing_arxiv_id":"2605.07922","citing_title":"Tree SAE: Learning Hierarchical Feature Structures in Sparse Autoencoders","ref_index":13,"is_internal_anchor":false},{"citing_arxiv_id":"2605.08012","citing_title":"Position: Mechanistic Interpretability Must Disclose Identification Assumptions for Causal Claims","ref_index":13,"is_internal_anchor":false}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/PWLSA3CRXYWS3I6LKJDKY7HZYL","json":"https://pith.science/pith/PWLSA3CRXYWS3I6LKJDKY7HZYL.json","graph_json":"https://pith.science/api/pith-number/PWLSA3CRXYWS3I6LKJDKY7HZYL/graph.json","events_json":"https://pith.science/api/pith-number/PWLSA3CRXYWS3I6LKJDKY7HZYL/events.json","paper":"https://pith.science/paper/PWLSA3CR"},"agent_actions":{"view_html":"https://pith.science/pith/PWLSA3CRXYWS3I6LKJDKY7HZYL","download_json":"https://pith.science/pith/PWLSA3CRXYWS3I6LKJDKY7HZYL.json","view_paper":"https://pith.science/paper/PWLSA3CR","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2502.04878&json=true","fetch_graph":"https://pith.science/api/pith-number/PWLSA3CRXYWS3I6LKJDKY7HZYL/graph.json","fetch_events":"https://pith.science/api/pith-number/PWLSA3CRXYWS3I6LKJDKY7HZYL/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/PWLSA3CRXYWS3I6LKJDKY7HZYL/action/timestamp_anchor","attest_storage":"https://pith.science/pith/PWLSA3CRXYWS3I6LKJDKY7HZYL/action/storage_attestation","attest_author":"https://pith.science/pith/PWLSA3CRXYWS3I6LKJDKY7HZYL/action/author_attestation","sign_citation":"https://pith.science/pith/PWLSA3CRXYWS3I6LKJDKY7HZYL/action/citation_signature","submit_replication":"https://pith.science/pith/PWLSA3CRXYWS3I6LKJDKY7HZYL/action/replication_record"}},"created_at":"2026-07-05T10:10:58.955049+00:00","updated_at":"2026-07-05T10:10:58.955049+00:00"}