{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2026:BOLR37ZKRVR2XXZQZD44HGFVZZ","short_pith_number":"pith:BOLR37ZK","schema_version":"1.0","canonical_sha256":"0b971dff2a8d63abdf30c8f9c398b5ce5a61795ae7e3d5587651a8476c0f9a29","source":{"kind":"arxiv","id":"2602.13483","version":2},"attestation_state":"computed","paper":{"title":"Finding Interpretable Prompt-Specific Circuits in Language Models","license":"http://creativecommons.org/licenses/by/4.0/","headline":"ACC++ extracts causal attention signals from language models in a single forward pass, revealing many are interpretable via natural language.","cross_cats":["cs.AI"],"primary_cat":"cs.LG","authors_text":"Azalea Rohr, Gabriel Franco, Lucas M. Tassis, Mark Crovella","submitted_at":"2026-02-13T21:41:17Z","abstract_excerpt":"Understanding the internal circuits that language models use to solve tasks remains a central challenge in mechanistic interpretability. A crucial part of finding circuits is understanding why each attention head attends where it does. To this end, we introduce ACC++, an improved circuit-tracing method based on the principle of attention-causal communication (ACC) [1], which identifies signals, i.e., contents of low dimensional subspaces that cause attention on a token pair. ACC++ extracts circuits from a single forward pass, without replacement models or patching. Circuits identified by ACC++"},"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":true,"formal_links_present":false},"canonical_record":{"source":{"id":"2602.13483","kind":"arxiv","version":2},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.LG","submitted_at":"2026-02-13T21:41:17Z","cross_cats_sorted":["cs.AI"],"title_canon_sha256":"86d076f178dc5ed41fb42c4f95499884d4ed051998d7844be5f96d493f8bee6b","abstract_canon_sha256":"29dc341a63e705f5542be6f9b1beb90f833899d66d98bae03ca3acaa88ddf930"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-05-17T23:39:16.172241Z","signature_b64":"w3NNWt0lg/4akZ5O7c5eOYeOMr/Rh7wzKR5d6lZOktCkaPn/+OXyPL20zA5vmC+fa3r9YOJ0zI9F5Z/NGuOBDA==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"0b971dff2a8d63abdf30c8f9c398b5ce5a61795ae7e3d5587651a8476c0f9a29","last_reissued_at":"2026-05-17T23:39:16.171443Z","signature_status":"signed_v1","first_computed_at":"2026-05-17T23:39:16.171443Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Finding Interpretable Prompt-Specific Circuits in Language Models","license":"http://creativecommons.org/licenses/by/4.0/","headline":"ACC++ extracts causal attention signals from language models in a single forward pass, revealing many are interpretable via natural language.","cross_cats":["cs.AI"],"primary_cat":"cs.LG","authors_text":"Azalea Rohr, Gabriel Franco, Lucas M. Tassis, Mark Crovella","submitted_at":"2026-02-13T21:41:17Z","abstract_excerpt":"Understanding the internal circuits that language models use to solve tasks remains a central challenge in mechanistic interpretability. A crucial part of finding circuits is understanding why each attention head attends where it does. To this end, we introduce ACC++, an improved circuit-tracing method based on the principle of attention-causal communication (ACC) [1], which identifies signals, i.e., contents of low dimensional subspaces that cause attention on a token pair. ACC++ extracts circuits from a single forward pass, without replacement models or patching. Circuits identified by ACC++"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"ACC++ extracts circuits from a single forward pass, without replacement models or patching. Circuits identified by ACC++ consist of components that are causal for the model's attention decisions, together with the low-dimensional signals used to communicate between them. Across multiple models, a substantial portion of ACC++ signals are interpretable: many signals admit a short natural-language description.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"That the low-dimensional subspaces identified by the attention-causal communication principle are in fact the causal signals driving attention decisions, and that the natural-language descriptions assigned to them reflect genuine model mechanisms rather than post-hoc interpretations.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"ACC++ traces prompt-specific circuits in language models from one forward pass by extracting interpretable low-dimensional causal signals, revealing clustered mechanisms for indirect object identification and language-specific signals in multilingual settings.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"ACC++ extracts causal attention signals from language models in a single forward pass, revealing many are interpretable via natural language.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"064ee08e1b484f1ebf162daa1bb73c89d5d5707a45d4e68f401420850e404808"},"source":{"id":"2602.13483","kind":"arxiv","version":2},"verdict":{"id":"348b5f11-9cd1-4430-813f-b5680bd5f004","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-15T22:03:52.950991Z","strongest_claim":"ACC++ extracts circuits from a single forward pass, without replacement models or patching. Circuits identified by ACC++ consist of components that are causal for the model's attention decisions, together with the low-dimensional signals used to communicate between them. Across multiple models, a substantial portion of ACC++ signals are interpretable: many signals admit a short natural-language description.","one_line_summary":"ACC++ traces prompt-specific circuits in language models from one forward pass by extracting interpretable low-dimensional causal signals, revealing clustered mechanisms for indirect object identification and language-specific signals in multilingual settings.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"That the low-dimensional subspaces identified by the attention-causal communication principle are in fact the causal signals driving attention decisions, and that the natural-language descriptions assigned to them reflect genuine model mechanisms rather than post-hoc interpretations.","pith_extraction_headline":"ACC++ extracts causal attention signals from language models in a single forward pass, revealing many are interpretable via natural language."},"references":{"count":18,"sample":[{"doi":"10.48550/arxiv.2307","year":2025,"title":"doi:10.48550/ARXIV.2307","work_id":"1b282611-e55d-4374-95d1-628b9e4bcd1f","ref_index":1,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":null,"title":"Set up the counterfactual.Choose an attention head and a destination–source pair, then decide whether we are searching fordestinationsignals (in the destination token) orsourcesignals (distributed acr","work_id":"9c99932e-180e-4878-b788-7649650dc9e2","ref_index":2,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":null,"title":"Enumerate candidate signals.Decompose the residual stream into outputs of upstream components, and project each component’s contribution onto the head’s singular-vector directions to form a set of can","work_id":"a44c09f5-d705-4486-b316-762de16fbff7","ref_index":3,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":null,"title":"Build a contribution table.Convert candidates into a contribution matrix whose rows correspond to candidate signals and whose columns correspond to source positions in the destination row of attention","work_id":"7041f06b-dbf6-4eb7-a5c3-2e1d03472a41","ref_index":4,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":null,"title":"Score candidates with attribution.Use Integrated Gradients to assign each candidate a fixed importance score for the attention weight on the chosen source token, while accounting for Softmax competiti","work_id":"fac3f535-87d1-4015-aa55-400822f413b1","ref_index":5,"cited_arxiv_id":"","is_internal_anchor":false}],"resolved_work":18,"snapshot_sha256":"ee4e0e0bb5dabc6b2aa4b7cbeaf7f41b3dfff7174bfbc1c49d5afd25f2688138","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":"2602.13483","created_at":"2026-05-17T23:39:16.171570+00:00"},{"alias_kind":"arxiv_version","alias_value":"2602.13483v2","created_at":"2026-05-17T23:39:16.171570+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2602.13483","created_at":"2026-05-17T23:39:16.171570+00:00"},{"alias_kind":"pith_short_12","alias_value":"BOLR37ZKRVR2","created_at":"2026-05-18T12:33:37.589309+00:00"},{"alias_kind":"pith_short_16","alias_value":"BOLR37ZKRVR2XXZQ","created_at":"2026-05-18T12:33:37.589309+00:00"},{"alias_kind":"pith_short_8","alias_value":"BOLR37ZK","created_at":"2026-05-18T12:33:37.589309+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":1,"internal_anchor_count":1,"sample":[{"citing_arxiv_id":"2605.09129","citing_title":"Data-driven Circuit Discovery for Interpretability of Language Models","ref_index":11,"is_internal_anchor":true}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/BOLR37ZKRVR2XXZQZD44HGFVZZ","json":"https://pith.science/pith/BOLR37ZKRVR2XXZQZD44HGFVZZ.json","graph_json":"https://pith.science/api/pith-number/BOLR37ZKRVR2XXZQZD44HGFVZZ/graph.json","events_json":"https://pith.science/api/pith-number/BOLR37ZKRVR2XXZQZD44HGFVZZ/events.json","paper":"https://pith.science/paper/BOLR37ZK"},"agent_actions":{"view_html":"https://pith.science/pith/BOLR37ZKRVR2XXZQZD44HGFVZZ","download_json":"https://pith.science/pith/BOLR37ZKRVR2XXZQZD44HGFVZZ.json","view_paper":"https://pith.science/paper/BOLR37ZK","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2602.13483&json=true","fetch_graph":"https://pith.science/api/pith-number/BOLR37ZKRVR2XXZQZD44HGFVZZ/graph.json","fetch_events":"https://pith.science/api/pith-number/BOLR37ZKRVR2XXZQZD44HGFVZZ/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/BOLR37ZKRVR2XXZQZD44HGFVZZ/action/timestamp_anchor","attest_storage":"https://pith.science/pith/BOLR37ZKRVR2XXZQZD44HGFVZZ/action/storage_attestation","attest_author":"https://pith.science/pith/BOLR37ZKRVR2XXZQZD44HGFVZZ/action/author_attestation","sign_citation":"https://pith.science/pith/BOLR37ZKRVR2XXZQZD44HGFVZZ/action/citation_signature","submit_replication":"https://pith.science/pith/BOLR37ZKRVR2XXZQZD44HGFVZZ/action/replication_record"}},"created_at":"2026-05-17T23:39:16.171570+00:00","updated_at":"2026-05-17T23:39:16.171570+00:00"}