{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2025:75TSUC5WDWTBDCNEQCD4OOX2AN","short_pith_number":"pith:75TSUC5W","schema_version":"1.0","canonical_sha256":"ff672a0bb61da61189a48087c73afa03547891dcb78399c2f7f33d23ed6f880d","source":{"kind":"arxiv","id":"2507.08473","version":1},"attestation_state":"computed","paper":{"title":"Evaluating SAE interpretability without explanations","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":[],"primary_cat":"cs.LG","authors_text":"Gon\\c{c}alo Paulo, Nora Belrose","submitted_at":"2025-07-11T10:31:53Z","abstract_excerpt":"Sparse autoencoders (SAEs) and transcoders have become important tools for machine learning interpretability. However, measuring how interpretable they are remains challenging, with weak consensus about which benchmarks to use. Most evaluation procedures start by producing a single-sentence explanation for each latent. These explanations are then evaluated based on how well they enable an LLM to predict the activation of a latent in new contexts. This method makes it difficult to disentangle the explanation generation and evaluation process from the actual interpretability of the latents disco"},"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":"2507.08473","kind":"arxiv","version":1},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.LG","submitted_at":"2025-07-11T10:31:53Z","cross_cats_sorted":[],"title_canon_sha256":"820e25e43b95a1564da66792652ca66a00c51bfae8a7dd02fe98ebb78c725383","abstract_canon_sha256":"b40ec80a8b2d25691565b4c2ed61d68bd9307605906b7d7571ef423a119110a8"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T11:35:40.536373Z","signature_b64":"cot2wiyOdUi/ei+f4vGAGsoo3qrL50JkHlgqbKmWEf0Gqte8Tl2eMtTSMuEaGyjYC91RP/yN2WE4BRWja3uzDw==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"ff672a0bb61da61189a48087c73afa03547891dcb78399c2f7f33d23ed6f880d","last_reissued_at":"2026-07-05T11:35:40.535754Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T11:35:40.535754Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Evaluating SAE interpretability without explanations","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":[],"primary_cat":"cs.LG","authors_text":"Gon\\c{c}alo Paulo, Nora Belrose","submitted_at":"2025-07-11T10:31:53Z","abstract_excerpt":"Sparse autoencoders (SAEs) and transcoders have become important tools for machine learning interpretability. However, measuring how interpretable they are remains challenging, with weak consensus about which benchmarks to use. Most evaluation procedures start by producing a single-sentence explanation for each latent. These explanations are then evaluated based on how well they enable an LLM to predict the activation of a latent in new contexts. This method makes it difficult to disentangle the explanation generation and evaluation process from the actual interpretability of the latents disco"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2507.08473","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/2507.08473/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":"2507.08473","created_at":"2026-07-05T11:35:40.535833+00:00"},{"alias_kind":"arxiv_version","alias_value":"2507.08473v1","created_at":"2026-07-05T11:35:40.535833+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2507.08473","created_at":"2026-07-05T11:35:40.535833+00:00"},{"alias_kind":"pith_short_12","alias_value":"75TSUC5WDWTB","created_at":"2026-07-05T11:35:40.535833+00:00"},{"alias_kind":"pith_short_16","alias_value":"75TSUC5WDWTBDCNE","created_at":"2026-07-05T11:35:40.535833+00:00"},{"alias_kind":"pith_short_8","alias_value":"75TSUC5W","created_at":"2026-07-05T11:35:40.535833+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":2,"internal_anchor_count":0,"sample":[{"citing_arxiv_id":"2605.28149","citing_title":"Sign-Aware Gated Sparse Autoencoders: Modeling Anticorrelated Features with Bi-Jump-ReLU Activations","ref_index":42,"is_internal_anchor":false},{"citing_arxiv_id":"2605.12874","citing_title":"Descriptive Collision in Sparse Autoencoder Auto-Interpretability: When One Explanation Describes Many Features","ref_index":22,"is_internal_anchor":false}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/75TSUC5WDWTBDCNEQCD4OOX2AN","json":"https://pith.science/pith/75TSUC5WDWTBDCNEQCD4OOX2AN.json","graph_json":"https://pith.science/api/pith-number/75TSUC5WDWTBDCNEQCD4OOX2AN/graph.json","events_json":"https://pith.science/api/pith-number/75TSUC5WDWTBDCNEQCD4OOX2AN/events.json","paper":"https://pith.science/paper/75TSUC5W"},"agent_actions":{"view_html":"https://pith.science/pith/75TSUC5WDWTBDCNEQCD4OOX2AN","download_json":"https://pith.science/pith/75TSUC5WDWTBDCNEQCD4OOX2AN.json","view_paper":"https://pith.science/paper/75TSUC5W","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2507.08473&json=true","fetch_graph":"https://pith.science/api/pith-number/75TSUC5WDWTBDCNEQCD4OOX2AN/graph.json","fetch_events":"https://pith.science/api/pith-number/75TSUC5WDWTBDCNEQCD4OOX2AN/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/75TSUC5WDWTBDCNEQCD4OOX2AN/action/timestamp_anchor","attest_storage":"https://pith.science/pith/75TSUC5WDWTBDCNEQCD4OOX2AN/action/storage_attestation","attest_author":"https://pith.science/pith/75TSUC5WDWTBDCNEQCD4OOX2AN/action/author_attestation","sign_citation":"https://pith.science/pith/75TSUC5WDWTBDCNEQCD4OOX2AN/action/citation_signature","submit_replication":"https://pith.science/pith/75TSUC5WDWTBDCNEQCD4OOX2AN/action/replication_record"}},"created_at":"2026-07-05T11:35:40.535833+00:00","updated_at":"2026-07-05T11:35:40.535833+00:00"}