{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2024:FS63F4S36EC7UPBRN5SMQ5FO5W","short_pith_number":"pith:FS63F4S3","schema_version":"1.0","canonical_sha256":"2cbdb2f25bf105fa3c316f64c874aeedad2def085a34ec37c1a214d3c4ea5f49","source":{"kind":"arxiv","id":"2412.03773","version":1},"attestation_state":"computed","paper":{"title":"Modular addition without black-boxes: Compressing explanations of MLPs that compute numerical integration","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.AI"],"primary_cat":"cs.LG","authors_text":"Chun Hei Yip, Jason Gross, Lawrence Chan, Rajashree Agrawal","submitted_at":"2024-12-04T23:29:07Z","abstract_excerpt":"The goal of mechanistic interpretability is discovering simpler, low-rank algorithms implemented by models. While we can compress activations into features, compressing nonlinear feature-maps -- like MLP layers -- is an open problem. In this work, we present the first case study in rigorously compressing nonlinear feature-maps, which are the leading asymptotic bottleneck to compressing small transformer models. We work in the classic setting of the modular addition models, and target a non-vacuous bound on the behaviour of the ReLU MLP in time linear in the parameter-count of the circuit. To s"},"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":"2412.03773","kind":"arxiv","version":1},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.LG","submitted_at":"2024-12-04T23:29:07Z","cross_cats_sorted":["cs.AI"],"title_canon_sha256":"48d0995b82e35840f558bc5b84556ad80976ca59f1734225fbe3301511cf9349","abstract_canon_sha256":"97a0214b276b84c50ee977bbceb5dcaf2f50b2baca87d13bf19f0d1f77614405"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T09:44:41.129202Z","signature_b64":"WNCXVbZIvmEvV85+glgteJiBXpK8+uak/emJjQL0ei+UasuSXF9BfAxclfcVOc8h0iaF0+DcIzlBAz0DMW/8DA==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"2cbdb2f25bf105fa3c316f64c874aeedad2def085a34ec37c1a214d3c4ea5f49","last_reissued_at":"2026-07-05T09:44:41.128747Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T09:44:41.128747Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Modular addition without black-boxes: Compressing explanations of MLPs that compute numerical integration","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.AI"],"primary_cat":"cs.LG","authors_text":"Chun Hei Yip, Jason Gross, Lawrence Chan, Rajashree Agrawal","submitted_at":"2024-12-04T23:29:07Z","abstract_excerpt":"The goal of mechanistic interpretability is discovering simpler, low-rank algorithms implemented by models. While we can compress activations into features, compressing nonlinear feature-maps -- like MLP layers -- is an open problem. In this work, we present the first case study in rigorously compressing nonlinear feature-maps, which are the leading asymptotic bottleneck to compressing small transformer models. We work in the classic setting of the modular addition models, and target a non-vacuous bound on the behaviour of the ReLU MLP in time linear in the parameter-count of the circuit. To s"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2412.03773","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/2412.03773/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":"2412.03773","created_at":"2026-07-05T09:44:41.128803+00:00"},{"alias_kind":"arxiv_version","alias_value":"2412.03773v1","created_at":"2026-07-05T09:44:41.128803+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2412.03773","created_at":"2026-07-05T09:44:41.128803+00:00"},{"alias_kind":"pith_short_12","alias_value":"FS63F4S36EC7","created_at":"2026-07-05T09:44:41.128803+00:00"},{"alias_kind":"pith_short_16","alias_value":"FS63F4S36EC7UPBR","created_at":"2026-07-05T09:44:41.128803+00:00"},{"alias_kind":"pith_short_8","alias_value":"FS63F4S3","created_at":"2026-07-05T09:44:41.128803+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":1,"internal_anchor_count":0,"sample":[{"citing_arxiv_id":"2606.09940","citing_title":"Interactions Between Crosscoder Features: A Compact Proofs Perspective","ref_index":39,"is_internal_anchor":false}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/FS63F4S36EC7UPBRN5SMQ5FO5W","json":"https://pith.science/pith/FS63F4S36EC7UPBRN5SMQ5FO5W.json","graph_json":"https://pith.science/api/pith-number/FS63F4S36EC7UPBRN5SMQ5FO5W/graph.json","events_json":"https://pith.science/api/pith-number/FS63F4S36EC7UPBRN5SMQ5FO5W/events.json","paper":"https://pith.science/paper/FS63F4S3"},"agent_actions":{"view_html":"https://pith.science/pith/FS63F4S36EC7UPBRN5SMQ5FO5W","download_json":"https://pith.science/pith/FS63F4S36EC7UPBRN5SMQ5FO5W.json","view_paper":"https://pith.science/paper/FS63F4S3","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2412.03773&json=true","fetch_graph":"https://pith.science/api/pith-number/FS63F4S36EC7UPBRN5SMQ5FO5W/graph.json","fetch_events":"https://pith.science/api/pith-number/FS63F4S36EC7UPBRN5SMQ5FO5W/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/FS63F4S36EC7UPBRN5SMQ5FO5W/action/timestamp_anchor","attest_storage":"https://pith.science/pith/FS63F4S36EC7UPBRN5SMQ5FO5W/action/storage_attestation","attest_author":"https://pith.science/pith/FS63F4S36EC7UPBRN5SMQ5FO5W/action/author_attestation","sign_citation":"https://pith.science/pith/FS63F4S36EC7UPBRN5SMQ5FO5W/action/citation_signature","submit_replication":"https://pith.science/pith/FS63F4S36EC7UPBRN5SMQ5FO5W/action/replication_record"}},"created_at":"2026-07-05T09:44:41.128803+00:00","updated_at":"2026-07-05T09:44:41.128803+00:00"}