{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2025:G2ZKLKWAO4TZROCU6ZGCWP7H5P","short_pith_number":"pith:G2ZKLKWA","schema_version":"1.0","canonical_sha256":"36b2a5aac0772798b854f64c2b3fe7ebda3f9bba525296330b61a12e93e0ada9","source":{"kind":"arxiv","id":"2505.24473","version":2},"attestation_state":"computed","paper":{"title":"Train One Sparse Autoencoder Across Multiple Sparsity Budgets to Preserve Interpretability and Accuracy","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.AI"],"primary_cat":"cs.LG","authors_text":"Daniil Gavrilov, Daniil Laptev, Gleb Gerasimov, Nikita Balagansky, Nikita Koryagin, Vadim Kurochkin, Yaroslav Aksenov","submitted_at":"2025-05-30T11:20:44Z","abstract_excerpt":"Sparse Autoencoders (SAEs) have proven to be powerful tools for interpreting neural networks by decomposing hidden representations into disentangled, interpretable features via sparsity constraints. However, conventional SAEs are constrained by the fixed sparsity level chosen during training; meeting different sparsity requirements therefore demands separate models and increases the computational footprint during both training and evaluation. We introduce a novel training objective, \\emph{HierarchicalTopK}, which trains a single SAE to optimise reconstructions across multiple sparsity levels s"},"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":"2505.24473","kind":"arxiv","version":2},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.LG","submitted_at":"2025-05-30T11:20:44Z","cross_cats_sorted":["cs.AI"],"title_canon_sha256":"05dd1d7162c2717bfd27d66978370278f38c844e96fad3bddb2c663431c870ef","abstract_canon_sha256":"eec0d7ea1d402a7432c90f4d564b973a9be1bcb5ed2e3dcfe6a23d5874346af7"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T11:16:35.764949Z","signature_b64":"hDgJGW2z+VS/7Pkjo44E79SEhYRxfXJXGbKenbXp/aI1Wf2LUnIA5BYQZlrseUwfPZCjlvBPYw0MDB0TBC87DQ==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"36b2a5aac0772798b854f64c2b3fe7ebda3f9bba525296330b61a12e93e0ada9","last_reissued_at":"2026-07-05T11:16:35.764329Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T11:16:35.764329Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Train One Sparse Autoencoder Across Multiple Sparsity Budgets to Preserve Interpretability and Accuracy","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.AI"],"primary_cat":"cs.LG","authors_text":"Daniil Gavrilov, Daniil Laptev, Gleb Gerasimov, Nikita Balagansky, Nikita Koryagin, Vadim Kurochkin, Yaroslav Aksenov","submitted_at":"2025-05-30T11:20:44Z","abstract_excerpt":"Sparse Autoencoders (SAEs) have proven to be powerful tools for interpreting neural networks by decomposing hidden representations into disentangled, interpretable features via sparsity constraints. However, conventional SAEs are constrained by the fixed sparsity level chosen during training; meeting different sparsity requirements therefore demands separate models and increases the computational footprint during both training and evaluation. We introduce a novel training objective, \\emph{HierarchicalTopK}, which trains a single SAE to optimise reconstructions across multiple sparsity levels s"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2505.24473","kind":"arxiv","version":2},"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/2505.24473/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":"2505.24473","created_at":"2026-07-05T11:16:35.764412+00:00"},{"alias_kind":"arxiv_version","alias_value":"2505.24473v2","created_at":"2026-07-05T11:16:35.764412+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2505.24473","created_at":"2026-07-05T11:16:35.764412+00:00"},{"alias_kind":"pith_short_12","alias_value":"G2ZKLKWAO4TZ","created_at":"2026-07-05T11:16:35.764412+00:00"},{"alias_kind":"pith_short_16","alias_value":"G2ZKLKWAO4TZROCU","created_at":"2026-07-05T11:16:35.764412+00:00"},{"alias_kind":"pith_short_8","alias_value":"G2ZKLKWA","created_at":"2026-07-05T11:16:35.764412+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":2,"internal_anchor_count":0,"sample":[{"citing_arxiv_id":"2606.18089","citing_title":"From Reasoning Traces to Reusable Modules: Understanding Compositional Generalization in Language Model Reasoning","ref_index":89,"is_internal_anchor":false},{"citing_arxiv_id":"2605.10536","citing_title":"HH-SAE: Discovering and Steering Hierarchical Knowledge of Complex Manifolds","ref_index":16,"is_internal_anchor":false}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/G2ZKLKWAO4TZROCU6ZGCWP7H5P","json":"https://pith.science/pith/G2ZKLKWAO4TZROCU6ZGCWP7H5P.json","graph_json":"https://pith.science/api/pith-number/G2ZKLKWAO4TZROCU6ZGCWP7H5P/graph.json","events_json":"https://pith.science/api/pith-number/G2ZKLKWAO4TZROCU6ZGCWP7H5P/events.json","paper":"https://pith.science/paper/G2ZKLKWA"},"agent_actions":{"view_html":"https://pith.science/pith/G2ZKLKWAO4TZROCU6ZGCWP7H5P","download_json":"https://pith.science/pith/G2ZKLKWAO4TZROCU6ZGCWP7H5P.json","view_paper":"https://pith.science/paper/G2ZKLKWA","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2505.24473&json=true","fetch_graph":"https://pith.science/api/pith-number/G2ZKLKWAO4TZROCU6ZGCWP7H5P/graph.json","fetch_events":"https://pith.science/api/pith-number/G2ZKLKWAO4TZROCU6ZGCWP7H5P/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/G2ZKLKWAO4TZROCU6ZGCWP7H5P/action/timestamp_anchor","attest_storage":"https://pith.science/pith/G2ZKLKWAO4TZROCU6ZGCWP7H5P/action/storage_attestation","attest_author":"https://pith.science/pith/G2ZKLKWAO4TZROCU6ZGCWP7H5P/action/author_attestation","sign_citation":"https://pith.science/pith/G2ZKLKWAO4TZROCU6ZGCWP7H5P/action/citation_signature","submit_replication":"https://pith.science/pith/G2ZKLKWAO4TZROCU6ZGCWP7H5P/action/replication_record"}},"created_at":"2026-07-05T11:16:35.764412+00:00","updated_at":"2026-07-05T11:16:35.764412+00:00"}