{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2024:DJMEJBMPCIGW7DYDVP3SJBX732","short_pith_number":"pith:DJMEJBMP","schema_version":"1.0","canonical_sha256":"1a5844858f120d6f8f03abf72486ffdeb32beeef86d0ddc6efbe78b30f57dcc0","source":{"kind":"arxiv","id":"2412.06410","version":1},"attestation_state":"computed","paper":{"title":"BatchTopK Sparse Autoencoders","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.AI","stat.ML"],"primary_cat":"cs.LG","authors_text":"Bart Bussmann, Neel Nanda, Patrick Leask","submitted_at":"2024-12-09T11:39:00Z","abstract_excerpt":"Sparse autoencoders (SAEs) have emerged as a powerful tool for interpreting language model activations by decomposing them into sparse, interpretable features. A popular approach is the TopK SAE, that uses a fixed number of the most active latents per sample to reconstruct the model activations. We introduce BatchTopK SAEs, a training method that improves upon TopK SAEs by relaxing the top-k constraint to the batch-level, allowing for a variable number of latents to be active per sample. As a result, BatchTopK adaptively allocates more or fewer latents depending on the sample, improving recons"},"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":"2412.06410","kind":"arxiv","version":1},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.LG","submitted_at":"2024-12-09T11:39:00Z","cross_cats_sorted":["cs.AI","stat.ML"],"title_canon_sha256":"15c34355bcda76a91839d3b2f683cb54adfa4a217b77530fc00ee6796b30d6da","abstract_canon_sha256":"8520f6b3e46c985151d0c1f374aea7cc820e9955ed00beeeb8d3267f012599ae"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T09:46:20.234759Z","signature_b64":"yNFEozyqVDaO5vZql9Dv3wglA/TRW1aaauLo4FABQOKbyUXFJ1zucLGL6bmRGxnIjUhCAYioNNG1OCpixZ0aAA==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"1a5844858f120d6f8f03abf72486ffdeb32beeef86d0ddc6efbe78b30f57dcc0","last_reissued_at":"2026-07-05T09:46:20.234353Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T09:46:20.234353Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"BatchTopK Sparse Autoencoders","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.AI","stat.ML"],"primary_cat":"cs.LG","authors_text":"Bart Bussmann, Neel Nanda, Patrick Leask","submitted_at":"2024-12-09T11:39:00Z","abstract_excerpt":"Sparse autoencoders (SAEs) have emerged as a powerful tool for interpreting language model activations by decomposing them into sparse, interpretable features. A popular approach is the TopK SAE, that uses a fixed number of the most active latents per sample to reconstruct the model activations. We introduce BatchTopK SAEs, a training method that improves upon TopK SAEs by relaxing the top-k constraint to the batch-level, allowing for a variable number of latents to be active per sample. As a result, BatchTopK adaptively allocates more or fewer latents depending on the sample, improving recons"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2412.06410","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/2412.06410/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":"2412.06410","created_at":"2026-07-05T09:46:20.234410+00:00"},{"alias_kind":"arxiv_version","alias_value":"2412.06410v1","created_at":"2026-07-05T09:46:20.234410+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2412.06410","created_at":"2026-07-05T09:46:20.234410+00:00"},{"alias_kind":"pith_short_12","alias_value":"DJMEJBMPCIGW","created_at":"2026-07-05T09:46:20.234410+00:00"},{"alias_kind":"pith_short_16","alias_value":"DJMEJBMPCIGW7DYD","created_at":"2026-07-05T09:46:20.234410+00:00"},{"alias_kind":"pith_short_8","alias_value":"DJMEJBMP","created_at":"2026-07-05T09:46:20.234410+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":34,"internal_anchor_count":1,"sample":[{"citing_arxiv_id":"2607.07316","citing_title":"Mechanistic Interpretability for Neural Networks: Circuits, Sparse Features and Symbolic Reasoning","ref_index":40,"is_internal_anchor":true},{"citing_arxiv_id":"2606.25657","citing_title":"Steering Vision-Language Models with Joint Sparse Autoencoders","ref_index":53,"is_internal_anchor":false},{"citing_arxiv_id":"2606.27321","citing_title":"Beyond the Hard Budget: Sparsity Regularizers for More Interpretable Top-k Sparse Autoencoders","ref_index":3,"is_internal_anchor":false},{"citing_arxiv_id":"2607.01799","citing_title":"Expander Sparse Autoencoders: Parameter-Efficient Dictionaries for Mechanistic Interpretability","ref_index":3,"is_internal_anchor":false},{"citing_arxiv_id":"2606.14990","citing_title":"Rational Sparse Autoencoder","ref_index":1,"is_internal_anchor":false},{"citing_arxiv_id":"2606.11722","citing_title":"ICA Lens: Interpreting Language Models Without Training Another Dictionary","ref_index":3,"is_internal_anchor":false},{"citing_arxiv_id":"2606.10080","citing_title":"VFUSE: Virulent Feature Understanding with Sparse autoEncoders","ref_index":13,"is_internal_anchor":false},{"citing_arxiv_id":"2606.09940","citing_title":"Interactions Between Crosscoder Features: A Compact Proofs Perspective","ref_index":4,"is_internal_anchor":false},{"citing_arxiv_id":"2607.00023","citing_title":"Aligning Sentence Embeddings to Human Concepts via Sparse Autoencoders","ref_index":1,"is_internal_anchor":false},{"citing_arxiv_id":"2606.07963","citing_title":"Shared Latent Structures Enable Unified Backdoor Detection and Mitigation in LLMs","ref_index":57,"is_internal_anchor":false},{"citing_arxiv_id":"2606.27440","citing_title":"PairSAE: Mechanistic Interpretability from Pair Representations in Protein Co-Folding","ref_index":11,"is_internal_anchor":false},{"citing_arxiv_id":"2606.28548","citing_title":"Turn-Averaged SAEs for Feature Discovery and Long-Context Attribution","ref_index":8,"is_internal_anchor":false},{"citing_arxiv_id":"2606.15054","citing_title":"Size Doesn't Matter: Cosine-Scored Sparse Autoencoders","ref_index":15,"is_internal_anchor":false},{"citing_arxiv_id":"2605.14694","citing_title":"The Rate-Distortion-Polysemanticity Tradeoff in SAEs","ref_index":4,"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":40,"is_internal_anchor":false},{"citing_arxiv_id":"2606.06333","citing_title":"Subspace-Aware Sparse Autoencoders for Effective Mechanistic Interpretability","ref_index":35,"is_internal_anchor":false},{"citing_arxiv_id":"2605.22679","citing_title":"Conceptualizing Embeddings: Sparse Disentanglement for Vision-Language Models","ref_index":31,"is_internal_anchor":false},{"citing_arxiv_id":"2602.11910","citing_title":"TADA! Tuning Audio Diffusion Models through Activation Steering","ref_index":5,"is_internal_anchor":false},{"citing_arxiv_id":"2605.12770","citing_title":"WriteSAE: Sparse Autoencoders for Recurrent State","ref_index":6,"is_internal_anchor":false},{"citing_arxiv_id":"2605.18229","citing_title":"Are Sparse Autoencoder Benchmarks Reliable?","ref_index":4,"is_internal_anchor":false},{"citing_arxiv_id":"2509.05291","citing_title":"Crosscoding Through Time: Tracking Emergence & Consolidation Of Linguistic Representations Throughout LLM Pretraining","ref_index":2,"is_internal_anchor":false},{"citing_arxiv_id":"2509.18127","citing_title":"Safe-SAIL: Towards a Fine-grained Safety Landscape of Large Language Models via Sparse Autoencoder Interpretation Framework","ref_index":7,"is_internal_anchor":false},{"citing_arxiv_id":"2601.14004","citing_title":"Locate, Steer, and Improve: A Practical Survey of Actionable Mechanistic Interpretability in Large Language Models","ref_index":27,"is_internal_anchor":false},{"citing_arxiv_id":"2605.12809","citing_title":"Correcting Influence: Unboxing LLM Outputs with Orthogonal Latent Spaces","ref_index":49,"is_internal_anchor":false},{"citing_arxiv_id":"2605.12122","citing_title":"Disentangled Sparse Representations for Concept-Separated Diffusion Unlearning","ref_index":2,"is_internal_anchor":false}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/DJMEJBMPCIGW7DYDVP3SJBX732","json":"https://pith.science/pith/DJMEJBMPCIGW7DYDVP3SJBX732.json","graph_json":"https://pith.science/api/pith-number/DJMEJBMPCIGW7DYDVP3SJBX732/graph.json","events_json":"https://pith.science/api/pith-number/DJMEJBMPCIGW7DYDVP3SJBX732/events.json","paper":"https://pith.science/paper/DJMEJBMP"},"agent_actions":{"view_html":"https://pith.science/pith/DJMEJBMPCIGW7DYDVP3SJBX732","download_json":"https://pith.science/pith/DJMEJBMPCIGW7DYDVP3SJBX732.json","view_paper":"https://pith.science/paper/DJMEJBMP","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2412.06410&json=true","fetch_graph":"https://pith.science/api/pith-number/DJMEJBMPCIGW7DYDVP3SJBX732/graph.json","fetch_events":"https://pith.science/api/pith-number/DJMEJBMPCIGW7DYDVP3SJBX732/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/DJMEJBMPCIGW7DYDVP3SJBX732/action/timestamp_anchor","attest_storage":"https://pith.science/pith/DJMEJBMPCIGW7DYDVP3SJBX732/action/storage_attestation","attest_author":"https://pith.science/pith/DJMEJBMPCIGW7DYDVP3SJBX732/action/author_attestation","sign_citation":"https://pith.science/pith/DJMEJBMPCIGW7DYDVP3SJBX732/action/citation_signature","submit_replication":"https://pith.science/pith/DJMEJBMPCIGW7DYDVP3SJBX732/action/replication_record"}},"created_at":"2026-07-05T09:46:20.234410+00:00","updated_at":"2026-07-05T09:46:20.234410+00:00"}