{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2024:QXWMAQZURQIIK5MLWJWKSA4UXL","short_pith_number":"pith:QXWMAQZU","schema_version":"1.0","canonical_sha256":"85ecc043348c1085758bb26ca90394baf62ecde2b4a65f0faa10053019f8335c","source":{"kind":"arxiv","id":"2404.16014","version":2},"attestation_state":"computed","paper":{"title":"Improving Dictionary Learning with Gated Sparse Autoencoders","license":"http://creativecommons.org/licenses/by/4.0/","headline":"Gated Sparse Autoencoders separate feature selection from magnitude estimation to eliminate L1-induced shrinkage in language model dictionary learning.","cross_cats":["cs.AI"],"primary_cat":"cs.LG","authors_text":"Arthur Conmy, J\\'anos Kram\\'ar, Lewis Smith, Neel Nanda, Rohin Shah, Senthooran Rajamanoharan, Tom Lieberum, Vikrant Varma","submitted_at":"2024-04-24T17:47:22Z","abstract_excerpt":"Recent work has found that sparse autoencoders (SAEs) are an effective technique for unsupervised discovery of interpretable features in language models' (LMs) activations, by finding sparse, linear reconstructions of LM activations. We introduce the Gated Sparse Autoencoder (Gated SAE), which achieves a Pareto improvement over training with prevailing methods. In SAEs, the L1 penalty used to encourage sparsity introduces many undesirable biases, such as shrinkage -- systematic underestimation of feature activations. The key insight of Gated SAEs is to separate the functionality of (a) determi"},"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":true,"formal_links_present":true},"canonical_record":{"source":{"id":"2404.16014","kind":"arxiv","version":2},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.LG","submitted_at":"2024-04-24T17:47:22Z","cross_cats_sorted":["cs.AI"],"title_canon_sha256":"de78f0873097f3b9f45e65322afe73347a1488e485186984f4ef162891cec806","abstract_canon_sha256":"a5a5978bca297540afaf137cdc1c11e59dd3aa7ff92132d2dba627675ae9dca9"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-05-17T23:38:13.271617Z","signature_b64":"0oL3berk1Lc7IeUz4AbT/k4YoQFcW7OWxAezVzKAI39MUS/NBJJaNAeq7sJxdMF/xgtLGq+pcbDHAeNEuuH/Cw==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"85ecc043348c1085758bb26ca90394baf62ecde2b4a65f0faa10053019f8335c","last_reissued_at":"2026-05-17T23:38:13.270899Z","signature_status":"signed_v1","first_computed_at":"2026-05-17T23:38:13.270899Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Improving Dictionary Learning with Gated Sparse Autoencoders","license":"http://creativecommons.org/licenses/by/4.0/","headline":"Gated Sparse Autoencoders separate feature selection from magnitude estimation to eliminate L1-induced shrinkage in language model dictionary learning.","cross_cats":["cs.AI"],"primary_cat":"cs.LG","authors_text":"Arthur Conmy, J\\'anos Kram\\'ar, Lewis Smith, Neel Nanda, Rohin Shah, Senthooran Rajamanoharan, Tom Lieberum, Vikrant Varma","submitted_at":"2024-04-24T17:47:22Z","abstract_excerpt":"Recent work has found that sparse autoencoders (SAEs) are an effective technique for unsupervised discovery of interpretable features in language models' (LMs) activations, by finding sparse, linear reconstructions of LM activations. We introduce the Gated Sparse Autoencoder (Gated SAE), which achieves a Pareto improvement over training with prevailing methods. In SAEs, the L1 penalty used to encourage sparsity introduces many undesirable biases, such as shrinkage -- systematic underestimation of feature activations. The key insight of Gated SAEs is to separate the functionality of (a) determi"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"Through training SAEs on LMs of up to 7B parameters we find that, in typical hyper-parameter ranges, Gated SAEs solve shrinkage, are similarly interpretable, and require half as many firing features to achieve comparable reconstruction fidelity.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"That restricting the L1 penalty to the gating branch does not introduce new biases or degrade feature quality in dimensions not measured by the reported reconstruction and interpretability metrics.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"Gated SAEs decouple which features to use from how large their activations should be, applying the L1 penalty only to selection and thereby eliminating shrinkage while halving the number of firing features needed for good fidelity.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"Gated Sparse Autoencoders separate feature selection from magnitude estimation to eliminate L1-induced shrinkage in language model dictionary learning.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"5f9f706a0b0938e3f3d8bc9ca1e3cb79457aa8c3c482ef27f9aa4ac918e18a20"},"source":{"id":"2404.16014","kind":"arxiv","version":2},"verdict":{"id":"2eaeba17-d035-40fe-846f-5ad164b529f3","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-17T19:36:14.845341Z","strongest_claim":"Through training SAEs on LMs of up to 7B parameters we find that, in typical hyper-parameter ranges, Gated SAEs solve shrinkage, are similarly interpretable, and require half as many firing features to achieve comparable reconstruction fidelity.","one_line_summary":"Gated SAEs decouple which features to use from how large their activations should be, applying the L1 penalty only to selection and thereby eliminating shrinkage while halving the number of firing features needed for good fidelity.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"That restricting the L1 penalty to the gating branch does not introduce new biases or degrade feature quality in dimensions not measured by the reported reconstruction and interpretability metrics.","pith_extraction_headline":"Gated Sparse Autoencoders separate feature selection from magnitude estimation to eliminate L1-induced shrinkage in language model dictionary learning."},"references":{"count":255,"sample":[{"doi":"10.1109/tsp.2006.881199","year":2006,"title":"M. Aharon, M. Elad, and A. Bruckstein. K-svd: An algorithm for designing overcomplete dictionaries for sparse representation. IEEE Transactions on Signal Processing, 54 0 (11): 0 4311--4322, 2006. doi","work_id":"3fa2e944-5f06-4fac-b63f-665b4e77cdfb","ref_index":1,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2024,"title":"Introducing the next generation of Claude","work_id":"e9acbb87-8254-4d24-9396-c241c784e85a","ref_index":2,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2024,"title":"J. Batson, B. Chen, A. Jones, A. Templeton, T. Conerly, J. Marcus, T. Henighan, N. L. Turner, and A. Pearce. Circuits Updates - March 2024 . Transformer Circuits Thread, 2024. URL https://transformer-","work_id":"9ca8ac9c-60e5-407c-8218-9753112be400","ref_index":3,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2013,"title":"Y. Bengio. Deep learning of representations: Looking forward, 2013","work_id":"2272cd0e-cdab-452c-a162-dc1d43e30767","ref_index":4,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2023,"title":"S. Biderman, H. Schoelkopf, Q. G. Anthony, H. Bradley, K. O’Brien, E. Hallahan, M. A. Khan, S. Purohit, U. S. Prashanth, E. Raff, et al. Pythia: A suite for analyzing large language models across trai","work_id":"c931f492-0319-4f63-a292-14ee570626a9","ref_index":5,"cited_arxiv_id":"","is_internal_anchor":false}],"resolved_work":255,"snapshot_sha256":"5943bf25d329c7adaa51ecd6c42589e405516a2b4ac7c0eff9d38ca097f6a88a","internal_anchors":5},"formal_canon":{"evidence_count":2,"snapshot_sha256":"96147c122c037e39ab96b8caa80eec45f8335c487dbf52c92c470271dba7619b"},"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":"2404.16014","created_at":"2026-05-17T23:38:13.271019+00:00"},{"alias_kind":"arxiv_version","alias_value":"2404.16014v2","created_at":"2026-05-17T23:38:13.271019+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2404.16014","created_at":"2026-05-17T23:38:13.271019+00:00"},{"alias_kind":"pith_short_12","alias_value":"QXWMAQZURQII","created_at":"2026-05-18T12:33:37.589309+00:00"},{"alias_kind":"pith_short_16","alias_value":"QXWMAQZURQIIK5ML","created_at":"2026-05-18T12:33:37.589309+00:00"},{"alias_kind":"pith_short_8","alias_value":"QXWMAQZU","created_at":"2026-05-18T12:33:37.589309+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":18,"internal_anchor_count":18,"sample":[{"citing_arxiv_id":"2412.14164","citing_title":"MetaMorph: Multimodal Understanding and Generation via Instruction Tuning","ref_index":233,"is_internal_anchor":true},{"citing_arxiv_id":"2601.14004","citing_title":"Locate, Steer, and Improve: A Practical Survey of Actionable Mechanistic Interpretability in Large Language Models","ref_index":252,"is_internal_anchor":true},{"citing_arxiv_id":"2605.12770","citing_title":"WriteSAE: Sparse Autoencoders for Recurrent State","ref_index":5,"is_internal_anchor":true},{"citing_arxiv_id":"2604.04946","citing_title":"Sparse Autoencoders as a Steering Basis for Phase Synchronization in Graph-Based CFD Surrogates","ref_index":10,"is_internal_anchor":true},{"citing_arxiv_id":"2605.12770","citing_title":"WriteSAE: Sparse Autoencoders for Recurrent State","ref_index":5,"is_internal_anchor":true},{"citing_arxiv_id":"2604.03436","citing_title":"MetaSAEs: Joint Training with a Decomposability Penalty Produces More Atomic Sparse Autoencoder Latents","ref_index":15,"is_internal_anchor":true},{"citing_arxiv_id":"2403.19647","citing_title":"Sparse Feature Circuits: Discovering and Editing Interpretable Causal Graphs in Language Models","ref_index":63,"is_internal_anchor":true},{"citing_arxiv_id":"2605.10933","citing_title":"DECO: Sparse Mixture-of-Experts with Dense-Comparable Performance on End-Side Devices","ref_index":163,"is_internal_anchor":true},{"citing_arxiv_id":"2406.04093","citing_title":"Scaling and evaluating sparse autoencoders","ref_index":52,"is_internal_anchor":true},{"citing_arxiv_id":"2605.09294","citing_title":"Towards Effective Theory of LLMs: A Representation Learning Approach","ref_index":47,"is_internal_anchor":true},{"citing_arxiv_id":"2605.10831","citing_title":"SLIM: Sparse Latent Steering for Interpretable and Property-Directed LLM-Based Molecular Editing","ref_index":9,"is_internal_anchor":true},{"citing_arxiv_id":"2605.10933","citing_title":"DECO: Sparse Mixture-of-Experts with Dense-Comparable Performance on End-Side Devices","ref_index":163,"is_internal_anchor":true},{"citing_arxiv_id":"2605.10536","citing_title":"HH-SAE: Discovering and Steering Hierarchical Knowledge of Complex Manifolds","ref_index":4,"is_internal_anchor":true},{"citing_arxiv_id":"2605.06494","citing_title":"From Token Lists to Graph Motifs: Weisfeiler-Lehman Analysis of Sparse Autoencoder Features","ref_index":5,"is_internal_anchor":true},{"citing_arxiv_id":"2604.08846","citing_title":"Dictionary-Aligned Concept Control for Safeguarding Multimodal LLMs","ref_index":85,"is_internal_anchor":true},{"citing_arxiv_id":"2605.07407","citing_title":"Emergent Symbolic Structure in Health Foundation Models: Extraction, Alignment, and Cross-Modal Transfer","ref_index":28,"is_internal_anchor":true},{"citing_arxiv_id":"2604.14925","citing_title":"Improving Sparse Autoencoder with Dynamic Attention","ref_index":57,"is_internal_anchor":true},{"citing_arxiv_id":"2605.05341","citing_title":"Feature Starvation as Geometric Instability in Sparse Autoencoders","ref_index":32,"is_internal_anchor":true}]},"formal_canon":{"evidence_count":2,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/QXWMAQZURQIIK5MLWJWKSA4UXL","json":"https://pith.science/pith/QXWMAQZURQIIK5MLWJWKSA4UXL.json","graph_json":"https://pith.science/api/pith-number/QXWMAQZURQIIK5MLWJWKSA4UXL/graph.json","events_json":"https://pith.science/api/pith-number/QXWMAQZURQIIK5MLWJWKSA4UXL/events.json","paper":"https://pith.science/paper/QXWMAQZU"},"agent_actions":{"view_html":"https://pith.science/pith/QXWMAQZURQIIK5MLWJWKSA4UXL","download_json":"https://pith.science/pith/QXWMAQZURQIIK5MLWJWKSA4UXL.json","view_paper":"https://pith.science/paper/QXWMAQZU","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2404.16014&json=true","fetch_graph":"https://pith.science/api/pith-number/QXWMAQZURQIIK5MLWJWKSA4UXL/graph.json","fetch_events":"https://pith.science/api/pith-number/QXWMAQZURQIIK5MLWJWKSA4UXL/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/QXWMAQZURQIIK5MLWJWKSA4UXL/action/timestamp_anchor","attest_storage":"https://pith.science/pith/QXWMAQZURQIIK5MLWJWKSA4UXL/action/storage_attestation","attest_author":"https://pith.science/pith/QXWMAQZURQIIK5MLWJWKSA4UXL/action/author_attestation","sign_citation":"https://pith.science/pith/QXWMAQZURQIIK5MLWJWKSA4UXL/action/citation_signature","submit_replication":"https://pith.science/pith/QXWMAQZURQIIK5MLWJWKSA4UXL/action/replication_record"}},"created_at":"2026-05-17T23:38:13.271019+00:00","updated_at":"2026-05-17T23:38:13.271019+00:00"}