{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2026:VEGH6MNY3GL3LTIQGU7MDD5LPI","short_pith_number":"pith:VEGH6MNY","schema_version":"1.0","canonical_sha256":"a90c7f31b8d997b5cd10353ec18fab7a3d17a0129a1f1946edb591c0e6e641d1","source":{"kind":"arxiv","id":"2601.05679","version":7},"attestation_state":"computed","paper":{"title":"Do Sparse Autoencoders Identify Reasoning Features in Language Models?","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":[],"primary_cat":"cs.LG","authors_text":"George Ma, Irene Y. Chen, Somayeh Sojoudi, Zhongyuan Liang","submitted_at":"2026-01-09T09:54:36Z","abstract_excerpt":"We study how reliably sparse autoencoders (SAEs) support claims about reasoning-related internal features in large language models. We first give a stylized analysis showing that sparsity-regularized decoding can preferentially retain stable low-dimensional correlates while suppressing high-dimensional within-behavior variation, motivating the possibility that contrastively selected \"reasoning\" features may concentrate on cue-like structure when such cues are coupled with reasoning traces. Building on this perspective, we propose a falsification-based evaluation framework that combines causal "},"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":"2601.05679","kind":"arxiv","version":7},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.LG","submitted_at":"2026-01-09T09:54:36Z","cross_cats_sorted":[],"title_canon_sha256":"209e5c216704f14d76bcedb11c317988df7f3d29d29a535ef40f41daa6d8a372","abstract_canon_sha256":"1b87fd1982d28e55d20b951bca480b72416c2a0aa8d7f9f90209635e7f11a927"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-05-20T00:03:01.822069Z","signature_b64":"E7hnAJ/YmKiYUasdiHF69cjPxdoLosWZci1K8/+ebhrbohFUtZ3swLnzb+gipeOubTcXiLH2OfQWfBn1sravDw==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"a90c7f31b8d997b5cd10353ec18fab7a3d17a0129a1f1946edb591c0e6e641d1","last_reissued_at":"2026-05-20T00:03:01.821187Z","signature_status":"signed_v1","first_computed_at":"2026-05-20T00:03:01.821187Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Do Sparse Autoencoders Identify Reasoning Features in Language Models?","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":[],"primary_cat":"cs.LG","authors_text":"George Ma, Irene Y. Chen, Somayeh Sojoudi, Zhongyuan Liang","submitted_at":"2026-01-09T09:54:36Z","abstract_excerpt":"We study how reliably sparse autoencoders (SAEs) support claims about reasoning-related internal features in large language models. We first give a stylized analysis showing that sparsity-regularized decoding can preferentially retain stable low-dimensional correlates while suppressing high-dimensional within-behavior variation, motivating the possibility that contrastively selected \"reasoning\" features may concentrate on cue-like structure when such cues are coupled with reasoning traces. Building on this perspective, we propose a falsification-based evaluation framework that combines causal "},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2601.05679","kind":"arxiv","version":7},"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/2601.05679/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":"2601.05679","created_at":"2026-05-20T00:03:01.821341+00:00"},{"alias_kind":"arxiv_version","alias_value":"2601.05679v7","created_at":"2026-05-20T00:03:01.821341+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2601.05679","created_at":"2026-05-20T00:03:01.821341+00:00"},{"alias_kind":"pith_short_12","alias_value":"VEGH6MNY3GL3","created_at":"2026-05-20T00:03:01.821341+00:00"},{"alias_kind":"pith_short_16","alias_value":"VEGH6MNY3GL3LTIQ","created_at":"2026-05-20T00:03:01.821341+00:00"},{"alias_kind":"pith_short_8","alias_value":"VEGH6MNY","created_at":"2026-05-20T00:03:01.821341+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":0,"internal_anchor_count":0,"sample":[]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/VEGH6MNY3GL3LTIQGU7MDD5LPI","json":"https://pith.science/pith/VEGH6MNY3GL3LTIQGU7MDD5LPI.json","graph_json":"https://pith.science/api/pith-number/VEGH6MNY3GL3LTIQGU7MDD5LPI/graph.json","events_json":"https://pith.science/api/pith-number/VEGH6MNY3GL3LTIQGU7MDD5LPI/events.json","paper":"https://pith.science/paper/VEGH6MNY"},"agent_actions":{"view_html":"https://pith.science/pith/VEGH6MNY3GL3LTIQGU7MDD5LPI","download_json":"https://pith.science/pith/VEGH6MNY3GL3LTIQGU7MDD5LPI.json","view_paper":"https://pith.science/paper/VEGH6MNY","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2601.05679&json=true","fetch_graph":"https://pith.science/api/pith-number/VEGH6MNY3GL3LTIQGU7MDD5LPI/graph.json","fetch_events":"https://pith.science/api/pith-number/VEGH6MNY3GL3LTIQGU7MDD5LPI/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/VEGH6MNY3GL3LTIQGU7MDD5LPI/action/timestamp_anchor","attest_storage":"https://pith.science/pith/VEGH6MNY3GL3LTIQGU7MDD5LPI/action/storage_attestation","attest_author":"https://pith.science/pith/VEGH6MNY3GL3LTIQGU7MDD5LPI/action/author_attestation","sign_citation":"https://pith.science/pith/VEGH6MNY3GL3LTIQGU7MDD5LPI/action/citation_signature","submit_replication":"https://pith.science/pith/VEGH6MNY3GL3LTIQGU7MDD5LPI/action/replication_record"}},"created_at":"2026-05-20T00:03:01.821341+00:00","updated_at":"2026-05-20T00:03:01.821341+00:00"}