{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2025:W6SWUHELBWKH6I3LQIPIXCICIE","short_pith_number":"pith:W6SWUHEL","schema_version":"1.0","canonical_sha256":"b7a56a1c8b0d947f236b821e8b8902412efc05c1890abe7420d8bcf4f1ce55f4","source":{"kind":"arxiv","id":"2504.13752","version":1},"attestation_state":"computed","paper":{"title":"Learning to Attribute with Attention","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"","cross_cats":["cs.CL"],"primary_cat":"cs.LG","authors_text":"Aleksander Madry, Benjamin Cohen-Wang, Yung-Sung Chuang","submitted_at":"2025-04-18T15:36:28Z","abstract_excerpt":"Given a sequence of tokens generated by a language model, we may want to identify the preceding tokens that influence the model to generate this sequence. Performing such token attribution is expensive; a common approach is to ablate preceding tokens and directly measure their effects. To reduce the cost of token attribution, we revisit attention weights as a heuristic for how a language model uses previous tokens. Naive approaches to attribute model behavior with attention (e.g., averaging attention weights across attention heads to estimate a token's influence) have been found to be unreliab"},"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":"2504.13752","kind":"arxiv","version":1},"metadata":{"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.LG","submitted_at":"2025-04-18T15:36:28Z","cross_cats_sorted":["cs.CL"],"title_canon_sha256":"a214c6b4d60338a1777e5b448b0ac6dc137e90cf0bb4d92446a0de47958a5d8e","abstract_canon_sha256":"280ecdb405b0bcf12bfca0a7697ef3ed3e1c1d32c3264b4fde3d4ebe9d0f634e"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T10:51:01.595873Z","signature_b64":"74Vm+DeqkO9LR17ux4JYp+2+dnb+chDfUCWlQYX294uVhBcl8Pz3ytUT+HDqQVS/F3xbxElQ/gF6qAETKVSfAA==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"b7a56a1c8b0d947f236b821e8b8902412efc05c1890abe7420d8bcf4f1ce55f4","last_reissued_at":"2026-07-05T10:51:01.595355Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T10:51:01.595355Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Learning to Attribute with Attention","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"","cross_cats":["cs.CL"],"primary_cat":"cs.LG","authors_text":"Aleksander Madry, Benjamin Cohen-Wang, Yung-Sung Chuang","submitted_at":"2025-04-18T15:36:28Z","abstract_excerpt":"Given a sequence of tokens generated by a language model, we may want to identify the preceding tokens that influence the model to generate this sequence. Performing such token attribution is expensive; a common approach is to ablate preceding tokens and directly measure their effects. To reduce the cost of token attribution, we revisit attention weights as a heuristic for how a language model uses previous tokens. Naive approaches to attribute model behavior with attention (e.g., averaging attention weights across attention heads to estimate a token's influence) have been found to be unreliab"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2504.13752","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/2504.13752/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":"2504.13752","created_at":"2026-07-05T10:51:01.595414+00:00"},{"alias_kind":"arxiv_version","alias_value":"2504.13752v1","created_at":"2026-07-05T10:51:01.595414+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2504.13752","created_at":"2026-07-05T10:51:01.595414+00:00"},{"alias_kind":"pith_short_12","alias_value":"W6SWUHELBWKH","created_at":"2026-07-05T10:51:01.595414+00:00"},{"alias_kind":"pith_short_16","alias_value":"W6SWUHELBWKH6I3L","created_at":"2026-07-05T10:51:01.595414+00:00"},{"alias_kind":"pith_short_8","alias_value":"W6SWUHEL","created_at":"2026-07-05T10:51:01.595414+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":2,"internal_anchor_count":0,"sample":[{"citing_arxiv_id":"2512.05958","citing_title":"MaxShapley: Towards Incentive-compatible Generative Search with Fair Context Attribution","ref_index":18,"is_internal_anchor":false},{"citing_arxiv_id":"2604.28157","citing_title":"FlashRT: Towards Computationally and Memory Efficient Red-Teaming for Prompt Injection and Knowledge Corruption","ref_index":48,"is_internal_anchor":false}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/W6SWUHELBWKH6I3LQIPIXCICIE","json":"https://pith.science/pith/W6SWUHELBWKH6I3LQIPIXCICIE.json","graph_json":"https://pith.science/api/pith-number/W6SWUHELBWKH6I3LQIPIXCICIE/graph.json","events_json":"https://pith.science/api/pith-number/W6SWUHELBWKH6I3LQIPIXCICIE/events.json","paper":"https://pith.science/paper/W6SWUHEL"},"agent_actions":{"view_html":"https://pith.science/pith/W6SWUHELBWKH6I3LQIPIXCICIE","download_json":"https://pith.science/pith/W6SWUHELBWKH6I3LQIPIXCICIE.json","view_paper":"https://pith.science/paper/W6SWUHEL","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2504.13752&json=true","fetch_graph":"https://pith.science/api/pith-number/W6SWUHELBWKH6I3LQIPIXCICIE/graph.json","fetch_events":"https://pith.science/api/pith-number/W6SWUHELBWKH6I3LQIPIXCICIE/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/W6SWUHELBWKH6I3LQIPIXCICIE/action/timestamp_anchor","attest_storage":"https://pith.science/pith/W6SWUHELBWKH6I3LQIPIXCICIE/action/storage_attestation","attest_author":"https://pith.science/pith/W6SWUHELBWKH6I3LQIPIXCICIE/action/author_attestation","sign_citation":"https://pith.science/pith/W6SWUHELBWKH6I3LQIPIXCICIE/action/citation_signature","submit_replication":"https://pith.science/pith/W6SWUHELBWKH6I3LQIPIXCICIE/action/replication_record"}},"created_at":"2026-07-05T10:51:01.595414+00:00","updated_at":"2026-07-05T10:51:01.595414+00:00"}