{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2026:LQB2CHW5G2XDMPAQB42YMKD4PV","short_pith_number":"pith:LQB2CHW5","schema_version":"1.0","canonical_sha256":"5c03a11edd36ae363c100f3586287c7d46e0d0062425052a2cab71466a630f29","source":{"kind":"arxiv","id":"2602.14687","version":2},"attestation_state":"computed","paper":{"title":"SynthSAEBench: Evaluating Sparse Autoencoders on Scalable Realistic Synthetic Data","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.AI"],"primary_cat":"cs.LG","authors_text":"Adri\\`a Garriga-Alonso, David Chanin","submitted_at":"2026-02-16T12:22:00Z","abstract_excerpt":"Improving Sparse Autoencoders (SAEs) requires benchmarks that can precisely validate architectural innovations. Current LLM-based SAE benchmarks are too noisy to differentiate architectural improvements, while commonly used synthetic-data experiments are too small-scale, unstandardized, and unrealistic to be meaningful. We introduce SynthSAEBench, a benchmark and toolkit for evaluating SAEs against large-scale synthetic data with realistic feature characteristics including correlation, hierarchy, and superposition, while providing ground-truth features and firings. SynthSAEBench acts as a cont"},"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":"2602.14687","kind":"arxiv","version":2},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.LG","submitted_at":"2026-02-16T12:22:00Z","cross_cats_sorted":["cs.AI"],"title_canon_sha256":"12b48d1cbd84f423e1324301fa15b50385d80105474cfcd183f235428fc5a7ef","abstract_canon_sha256":"2c8e0373d4ea6fea7c52848ca669d45887a3eec65be048c5a4bdfda7ef67146d"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-14T01:20:54.541390Z","signature_b64":"b5oCnNYoCEnY2iNhIb3DYX4eugAvuxga6EzJfNTMKhBT/JZy8uuJ8WiNXhHp0sjq1W00MV2J0VoNqAdyax7wDw==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"5c03a11edd36ae363c100f3586287c7d46e0d0062425052a2cab71466a630f29","last_reissued_at":"2026-07-14T01:20:54.540464Z","signature_status":"signed_v1","first_computed_at":"2026-07-14T01:20:54.540464Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"SynthSAEBench: Evaluating Sparse Autoencoders on Scalable Realistic Synthetic Data","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.AI"],"primary_cat":"cs.LG","authors_text":"Adri\\`a Garriga-Alonso, David Chanin","submitted_at":"2026-02-16T12:22:00Z","abstract_excerpt":"Improving Sparse Autoencoders (SAEs) requires benchmarks that can precisely validate architectural innovations. Current LLM-based SAE benchmarks are too noisy to differentiate architectural improvements, while commonly used synthetic-data experiments are too small-scale, unstandardized, and unrealistic to be meaningful. We introduce SynthSAEBench, a benchmark and toolkit for evaluating SAEs against large-scale synthetic data with realistic feature characteristics including correlation, hierarchy, and superposition, while providing ground-truth features and firings. SynthSAEBench acts as a cont"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2602.14687","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/2602.14687/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":"2602.14687","created_at":"2026-07-14T01:20:54.540895+00:00"},{"alias_kind":"arxiv_version","alias_value":"2602.14687v2","created_at":"2026-07-14T01:20:54.540895+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2602.14687","created_at":"2026-07-14T01:20:54.540895+00:00"},{"alias_kind":"pith_short_12","alias_value":"LQB2CHW5G2XD","created_at":"2026-07-14T01:20:54.540895+00:00"},{"alias_kind":"pith_short_16","alias_value":"LQB2CHW5G2XDMPAQ","created_at":"2026-07-14T01:20:54.540895+00:00"},{"alias_kind":"pith_short_8","alias_value":"LQB2CHW5","created_at":"2026-07-14T01:20:54.540895+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":4,"internal_anchor_count":4,"sample":[{"citing_arxiv_id":"2606.26620","citing_title":"Discovering Millions of Interpretable Features with Sparse Autoencoders","ref_index":6,"is_internal_anchor":true},{"citing_arxiv_id":"2606.20347","citing_title":"Critical Percolation as a Synthetic Data Model for Interpretability","ref_index":14,"is_internal_anchor":true},{"citing_arxiv_id":"2605.14694","citing_title":"The Rate-Distortion-Polysemanticity Tradeoff in SAEs","ref_index":6,"is_internal_anchor":true},{"citing_arxiv_id":"2605.18229","citing_title":"Are Sparse Autoencoder Benchmarks Reliable?","ref_index":7,"is_internal_anchor":true}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/LQB2CHW5G2XDMPAQB42YMKD4PV","json":"https://pith.science/pith/LQB2CHW5G2XDMPAQB42YMKD4PV.json","graph_json":"https://pith.science/api/pith-number/LQB2CHW5G2XDMPAQB42YMKD4PV/graph.json","events_json":"https://pith.science/api/pith-number/LQB2CHW5G2XDMPAQB42YMKD4PV/events.json","paper":"https://pith.science/paper/LQB2CHW5"},"agent_actions":{"view_html":"https://pith.science/pith/LQB2CHW5G2XDMPAQB42YMKD4PV","download_json":"https://pith.science/pith/LQB2CHW5G2XDMPAQB42YMKD4PV.json","view_paper":"https://pith.science/paper/LQB2CHW5","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2602.14687&json=true","fetch_graph":"https://pith.science/api/pith-number/LQB2CHW5G2XDMPAQB42YMKD4PV/graph.json","fetch_events":"https://pith.science/api/pith-number/LQB2CHW5G2XDMPAQB42YMKD4PV/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/LQB2CHW5G2XDMPAQB42YMKD4PV/action/timestamp_anchor","attest_storage":"https://pith.science/pith/LQB2CHW5G2XDMPAQB42YMKD4PV/action/storage_attestation","attest_author":"https://pith.science/pith/LQB2CHW5G2XDMPAQB42YMKD4PV/action/author_attestation","sign_citation":"https://pith.science/pith/LQB2CHW5G2XDMPAQB42YMKD4PV/action/citation_signature","submit_replication":"https://pith.science/pith/LQB2CHW5G2XDMPAQB42YMKD4PV/action/replication_record"}},"created_at":"2026-07-14T01:20:54.540895+00:00","updated_at":"2026-07-14T01:20:54.540895+00:00"}