{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2025:VE7YB4QMSVTLLRI4O7IOAVPM2W","short_pith_number":"pith:VE7YB4QM","schema_version":"1.0","canonical_sha256":"a93f80f20c9566b5c51c77d0e055ecd5a80134d9af7070cca60272de5c8a8e13","source":{"kind":"arxiv","id":"2509.02565","version":2},"attestation_state":"computed","paper":{"title":"Understanding sparse autoencoder scaling in the presence of feature manifolds","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":[],"primary_cat":"cs.LG","authors_text":"Eric J. Michaud, Liv Gorton, Tom McGrath","submitted_at":"2025-09-02T17:59:50Z","abstract_excerpt":"Sparse autoencoders (SAEs) model the activations of a neural network as linear combinations of sparsely occurring directions of variation (latents). The ability of SAEs to reconstruct activations follows scaling laws w.r.t. the number of latents. In this work, we adapt a capacity-allocation model from the neural scaling literature (Brill, 2024) to understand SAE scaling, and in particular, to understand how \"feature manifolds\" (multi-dimensional features) influence scaling behavior. Consistent with prior work, the model recovers distinct scaling regimes. Notably, in one regime, feature manifol"},"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":"2509.02565","kind":"arxiv","version":2},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.LG","submitted_at":"2025-09-02T17:59:50Z","cross_cats_sorted":[],"title_canon_sha256":"f2ba6397cee0b3130c25904fde99c0c3b84f9e8e1f4bc6fba9673628e25445dc","abstract_canon_sha256":"75b98eda3be66944f5bc40ce297f646b56a29ee40330fbcabee98d7724e58afc"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T12:05:04.228837Z","signature_b64":"WDQ7laX903xAZVLweBxmXlPLCludnSEmlI32PO6+qNw+BfSxRMofiJwLLmrxrdBDYisvBpa3dcUfyofcLQeaBA==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"a93f80f20c9566b5c51c77d0e055ecd5a80134d9af7070cca60272de5c8a8e13","last_reissued_at":"2026-07-05T12:05:04.228320Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T12:05:04.228320Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Understanding sparse autoencoder scaling in the presence of feature manifolds","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":[],"primary_cat":"cs.LG","authors_text":"Eric J. Michaud, Liv Gorton, Tom McGrath","submitted_at":"2025-09-02T17:59:50Z","abstract_excerpt":"Sparse autoencoders (SAEs) model the activations of a neural network as linear combinations of sparsely occurring directions of variation (latents). The ability of SAEs to reconstruct activations follows scaling laws w.r.t. the number of latents. In this work, we adapt a capacity-allocation model from the neural scaling literature (Brill, 2024) to understand SAE scaling, and in particular, to understand how \"feature manifolds\" (multi-dimensional features) influence scaling behavior. Consistent with prior work, the model recovers distinct scaling regimes. Notably, in one regime, feature manifol"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2509.02565","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/2509.02565/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":"2509.02565","created_at":"2026-07-05T12:05:04.228387+00:00"},{"alias_kind":"arxiv_version","alias_value":"2509.02565v2","created_at":"2026-07-05T12:05:04.228387+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2509.02565","created_at":"2026-07-05T12:05:04.228387+00:00"},{"alias_kind":"pith_short_12","alias_value":"VE7YB4QMSVTL","created_at":"2026-07-05T12:05:04.228387+00:00"},{"alias_kind":"pith_short_16","alias_value":"VE7YB4QMSVTLLRI4","created_at":"2026-07-05T12:05:04.228387+00:00"},{"alias_kind":"pith_short_8","alias_value":"VE7YB4QM","created_at":"2026-07-05T12:05:04.228387+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":5,"internal_anchor_count":0,"sample":[{"citing_arxiv_id":"2606.20347","citing_title":"Critical Percolation as a Synthetic Data Model for Interpretability","ref_index":37,"is_internal_anchor":false},{"citing_arxiv_id":"2606.07007","citing_title":"A Geometric View for Understanding Concept Learning and Neuron Interpretation in Sparse Autoencoders","ref_index":18,"is_internal_anchor":false},{"citing_arxiv_id":"2605.29548","citing_title":"Why Larger Models Learn More: Effects of Capacity, Interference, and Rare-Task Retention","ref_index":137,"is_internal_anchor":false},{"citing_arxiv_id":"2606.06333","citing_title":"Subspace-Aware Sparse Autoencoders for Effective Mechanistic Interpretability","ref_index":37,"is_internal_anchor":false},{"citing_arxiv_id":"2605.01192","citing_title":"Linear-Readout Floors and Threshold Recovery in Computation in Superposition","ref_index":12,"is_internal_anchor":false}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/VE7YB4QMSVTLLRI4O7IOAVPM2W","json":"https://pith.science/pith/VE7YB4QMSVTLLRI4O7IOAVPM2W.json","graph_json":"https://pith.science/api/pith-number/VE7YB4QMSVTLLRI4O7IOAVPM2W/graph.json","events_json":"https://pith.science/api/pith-number/VE7YB4QMSVTLLRI4O7IOAVPM2W/events.json","paper":"https://pith.science/paper/VE7YB4QM"},"agent_actions":{"view_html":"https://pith.science/pith/VE7YB4QMSVTLLRI4O7IOAVPM2W","download_json":"https://pith.science/pith/VE7YB4QMSVTLLRI4O7IOAVPM2W.json","view_paper":"https://pith.science/paper/VE7YB4QM","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2509.02565&json=true","fetch_graph":"https://pith.science/api/pith-number/VE7YB4QMSVTLLRI4O7IOAVPM2W/graph.json","fetch_events":"https://pith.science/api/pith-number/VE7YB4QMSVTLLRI4O7IOAVPM2W/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/VE7YB4QMSVTLLRI4O7IOAVPM2W/action/timestamp_anchor","attest_storage":"https://pith.science/pith/VE7YB4QMSVTLLRI4O7IOAVPM2W/action/storage_attestation","attest_author":"https://pith.science/pith/VE7YB4QMSVTLLRI4O7IOAVPM2W/action/author_attestation","sign_citation":"https://pith.science/pith/VE7YB4QMSVTLLRI4O7IOAVPM2W/action/citation_signature","submit_replication":"https://pith.science/pith/VE7YB4QMSVTLLRI4O7IOAVPM2W/action/replication_record"}},"created_at":"2026-07-05T12:05:04.228387+00:00","updated_at":"2026-07-05T12:05:04.228387+00:00"}