{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2024:PZJ6SQLYJSZAQPBFYZHS2WVGQO","short_pith_number":"pith:PZJ6SQLY","schema_version":"1.0","canonical_sha256":"7e53e941784cb2083c25c64f2d5aa683892335135c2ec4402821756985b33f8e","source":{"kind":"arxiv","id":"2409.15019","version":2},"attestation_state":"computed","paper":{"title":"Evaluating Synthetic Activations composed of SAE Latents in GPT-2","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"","cross_cats":[],"primary_cat":"cs.LG","authors_text":"Chatrik Singh Mangat, Giorgi Giglemiani, Jett Janiak, Nora Petrova, Stefan Heimersheim","submitted_at":"2024-09-23T13:46:38Z","abstract_excerpt":"Sparse Auto-Encoders (SAEs) are commonly employed in mechanistic interpretability to decompose the residual stream into monosemantic SAE latents. Recent work demonstrates that perturbing a model's activations at an early layer results in a step-function-like change in the model's final layer activations. Furthermore, the model's sensitivity to this perturbation differs between model-generated (real) activations and random activations. In our study, we assess model sensitivity in order to compare real activations to synthetic activations composed of SAE latents. Our findings indicate that synth"},"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":"2409.15019","kind":"arxiv","version":2},"metadata":{"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.LG","submitted_at":"2024-09-23T13:46:38Z","cross_cats_sorted":[],"title_canon_sha256":"aedd3d98ec29943b85e8038ee5bb1e68b999e93b6ae10a58433147d40f171b16","abstract_canon_sha256":"8a0ceeceb89bb1e6d69b6b30dd2cba6b9b1867554d398baff52e9887c8618d87"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T09:36:48.311575Z","signature_b64":"yE3sZd8bqOUVinUluKMnF3+I13mu6xeCYVPsiBQbU+bU85RoEPSBXVxhocTjhqWNFnj0d1t7B/pOW9kfrqYqAQ==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"7e53e941784cb2083c25c64f2d5aa683892335135c2ec4402821756985b33f8e","last_reissued_at":"2026-07-05T09:36:48.311090Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T09:36:48.311090Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Evaluating Synthetic Activations composed of SAE Latents in GPT-2","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"","cross_cats":[],"primary_cat":"cs.LG","authors_text":"Chatrik Singh Mangat, Giorgi Giglemiani, Jett Janiak, Nora Petrova, Stefan Heimersheim","submitted_at":"2024-09-23T13:46:38Z","abstract_excerpt":"Sparse Auto-Encoders (SAEs) are commonly employed in mechanistic interpretability to decompose the residual stream into monosemantic SAE latents. Recent work demonstrates that perturbing a model's activations at an early layer results in a step-function-like change in the model's final layer activations. Furthermore, the model's sensitivity to this perturbation differs between model-generated (real) activations and random activations. In our study, we assess model sensitivity in order to compare real activations to synthetic activations composed of SAE latents. Our findings indicate that synth"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2409.15019","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/2409.15019/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":"2409.15019","created_at":"2026-07-05T09:36:48.311146+00:00"},{"alias_kind":"arxiv_version","alias_value":"2409.15019v2","created_at":"2026-07-05T09:36:48.311146+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2409.15019","created_at":"2026-07-05T09:36:48.311146+00:00"},{"alias_kind":"pith_short_12","alias_value":"PZJ6SQLYJSZA","created_at":"2026-07-05T09:36:48.311146+00:00"},{"alias_kind":"pith_short_16","alias_value":"PZJ6SQLYJSZAQPBF","created_at":"2026-07-05T09:36:48.311146+00:00"},{"alias_kind":"pith_short_8","alias_value":"PZJ6SQLY","created_at":"2026-07-05T09:36:48.311146+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":1,"internal_anchor_count":1,"sample":[{"citing_arxiv_id":"2411.10397","citing_title":"Features that Make a Difference: Leveraging Gradients for Improved Dictionary Learning","ref_index":8,"is_internal_anchor":true}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/PZJ6SQLYJSZAQPBFYZHS2WVGQO","json":"https://pith.science/pith/PZJ6SQLYJSZAQPBFYZHS2WVGQO.json","graph_json":"https://pith.science/api/pith-number/PZJ6SQLYJSZAQPBFYZHS2WVGQO/graph.json","events_json":"https://pith.science/api/pith-number/PZJ6SQLYJSZAQPBFYZHS2WVGQO/events.json","paper":"https://pith.science/paper/PZJ6SQLY"},"agent_actions":{"view_html":"https://pith.science/pith/PZJ6SQLYJSZAQPBFYZHS2WVGQO","download_json":"https://pith.science/pith/PZJ6SQLYJSZAQPBFYZHS2WVGQO.json","view_paper":"https://pith.science/paper/PZJ6SQLY","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2409.15019&json=true","fetch_graph":"https://pith.science/api/pith-number/PZJ6SQLYJSZAQPBFYZHS2WVGQO/graph.json","fetch_events":"https://pith.science/api/pith-number/PZJ6SQLYJSZAQPBFYZHS2WVGQO/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/PZJ6SQLYJSZAQPBFYZHS2WVGQO/action/timestamp_anchor","attest_storage":"https://pith.science/pith/PZJ6SQLYJSZAQPBFYZHS2WVGQO/action/storage_attestation","attest_author":"https://pith.science/pith/PZJ6SQLYJSZAQPBFYZHS2WVGQO/action/author_attestation","sign_citation":"https://pith.science/pith/PZJ6SQLYJSZAQPBFYZHS2WVGQO/action/citation_signature","submit_replication":"https://pith.science/pith/PZJ6SQLYJSZAQPBFYZHS2WVGQO/action/replication_record"}},"created_at":"2026-07-05T09:36:48.311146+00:00","updated_at":"2026-07-05T09:36:48.311146+00:00"}