{"bundle_type":"pith_open_graph_bundle","bundle_version":"1.0","pith_number":"pith:2022:B7JM6H5UFDNTCIE5PSOVEXGQZB","short_pith_number":"pith:B7JM6H5U","canonical_record":{"source":{"id":"2211.00593","kind":"arxiv","version":1},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.LG","submitted_at":"2022-11-01T17:08:44Z","cross_cats_sorted":["cs.AI","cs.CL"],"title_canon_sha256":"5e96390378e8e0b98d86343f96a46f7677ac0ba7a59d5484d45b812fd5baea4b","abstract_canon_sha256":"63e3a75ec03b2c90703e86ed86c17d888e130ec03032fbf52ecd408fd9e5054e"},"schema_version":"1.0"},"canonical_sha256":"0fd2cf1fb428db31209d7c9d525cd0c870d4375341962ea0637c80732f2b4942","source":{"kind":"arxiv","id":"2211.00593","version":1},"source_aliases":[{"alias_kind":"arxiv","alias_value":"2211.00593","created_at":"2026-07-05T05:12:24Z"},{"alias_kind":"arxiv_version","alias_value":"2211.00593v1","created_at":"2026-07-05T05:12:24Z"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2211.00593","created_at":"2026-07-05T05:12:24Z"},{"alias_kind":"pith_short_12","alias_value":"B7JM6H5UFDNT","created_at":"2026-07-05T05:12:24Z"},{"alias_kind":"pith_short_16","alias_value":"B7JM6H5UFDNTCIE5","created_at":"2026-07-05T05:12:24Z"},{"alias_kind":"pith_short_8","alias_value":"B7JM6H5U","created_at":"2026-07-05T05:12:24Z"}],"events":[{"event_type":"record_created","subject_pith_number":"pith:2022:B7JM6H5UFDNTCIE5PSOVEXGQZB","target":"record","payload":{"canonical_record":{"source":{"id":"2211.00593","kind":"arxiv","version":1},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.LG","submitted_at":"2022-11-01T17:08:44Z","cross_cats_sorted":["cs.AI","cs.CL"],"title_canon_sha256":"5e96390378e8e0b98d86343f96a46f7677ac0ba7a59d5484d45b812fd5baea4b","abstract_canon_sha256":"63e3a75ec03b2c90703e86ed86c17d888e130ec03032fbf52ecd408fd9e5054e"},"schema_version":"1.0"},"canonical_sha256":"0fd2cf1fb428db31209d7c9d525cd0c870d4375341962ea0637c80732f2b4942","receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T05:12:24.537999Z","signature_b64":"weo05JfLya31WMljNrTtqDDjCAz5j28ntQ3WBB0lKe1debLitHzDWl+qxYMOcH8/04uNrbvHpnM59TeH53U6Cg==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"0fd2cf1fb428db31209d7c9d525cd0c870d4375341962ea0637c80732f2b4942","last_reissued_at":"2026-07-05T05:12:24.537581Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T05:12:24.537581Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"source_kind":"arxiv","source_id":"2211.00593","source_version":1,"attestation_state":"computed"},"signer":{"signer_id":"pith.science","signer_type":"pith_registry","key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"created_at":"2026-07-05T05:12:24Z","supersedes":[],"prev_event":null,"signature":{"signature_status":"signed_v1","algorithm":"ed25519","key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54","signature_b64":"msjKkv9jJs1mwkIwI6kFPF0mxcCLWL3lumiQ35R3g6MCcLs8YtyGEXsl0Iibx6yhXzLn9Rv9VHlA55LXYcKQAQ==","signed_message":"open_graph_event_sha256_bytes","signed_at":"2026-08-10T22:16:55.491986Z"},"content_sha256":"51ac6fa85937f515f21a435aba7e6d59b0cfc2ece93fbfbc71695832cd6d624c","schema_version":"1.0","event_id":"sha256:51ac6fa85937f515f21a435aba7e6d59b0cfc2ece93fbfbc71695832cd6d624c"},{"event_type":"graph_snapshot","subject_pith_number":"pith:2022:B7JM6H5UFDNTCIE5PSOVEXGQZB","target":"graph","payload":{"graph_snapshot":{"paper":{"title":"Interpretability in the Wild: a Circuit for Indirect Object Identification in GPT-2 small","license":"http://creativecommons.org/licenses/by/4.0/","headline":"GPT-2 small solves indirect object identification using a circuit of 26 attention heads in seven classes.","cross_cats":["cs.AI","cs.CL"],"primary_cat":"cs.LG","authors_text":"Alexandre Variengien, Arthur Conmy, Buck Shlegeris, Jacob Steinhardt, Kevin Wang","submitted_at":"2022-11-01T17:08:44Z","abstract_excerpt":"Research in mechanistic interpretability seeks to explain behaviors of machine learning models in terms of their internal components. However, most previous work either focuses on simple behaviors in small models, or describes complicated behaviors in larger models with broad strokes. In this work, we bridge this gap by presenting an explanation for how GPT-2 small performs a natural language task called indirect object identification (IOI). Our explanation encompasses 26 attention heads grouped into 7 main classes, which we discovered using a combination of interpretability approaches relying"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"Our explanation encompasses 26 attention heads grouped into 7 main classes, which we discovered using a combination of interpretability approaches relying on causal interventions. To our knowledge, this investigation is the largest end-to-end attempt at reverse-engineering a natural behavior 'in the wild' in a language model.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"That the three quantitative criteria (faithfulness, completeness, minimality) are sufficient to certify that the identified circuit is the complete and minimal explanation rather than one of several possible circuits that achieve similar performance.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"GPT-2 small solves indirect object identification via a circuit of 26 attention heads organized into seven functional classes discovered through causal interventions.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"GPT-2 small solves indirect object identification using a circuit of 26 attention heads in seven classes.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"83622a4e2142242358d616bb24d09074ac2054bf96f77e81a564b1cd170bbc9f"},"source":{"id":"2211.00593","kind":"arxiv","version":1},"verdict":{"id":"cfd4a671-6f3d-4097-9a96-a79116a5f62e","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-13T17:08:49.885716Z","strongest_claim":"Our explanation encompasses 26 attention heads grouped into 7 main classes, which we discovered using a combination of interpretability approaches relying on causal interventions. To our knowledge, this investigation is the largest end-to-end attempt at reverse-engineering a natural behavior 'in the wild' in a language model.","one_line_summary":"GPT-2 small solves indirect object identification via a circuit of 26 attention heads organized into seven functional classes discovered through causal interventions.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"That the three quantitative criteria (faithfulness, completeness, minimality) are sufficient to certify that the identified circuit is the complete and minimal explanation rather than one of several possible circuits that achieve similar performance.","pith_extraction_headline":"GPT-2 small solves indirect object identification using a circuit of 26 attention heads in seven classes."},"integrity":{"clean":true,"summary":{"advisory":0,"critical":0,"by_detector":{},"informational":0},"endpoint":"/pith/2211.00593/integrity.json","findings":[],"available":true,"detectors_run":[],"snapshot_sha256":"c28c3603d3b5d939e8dc4c7e95fa8dfce3d595e45f758748cecf8e644a296938"},"references":{"count":66,"sample":[{"doi":"","year":1901,"title":"Language models are few-shot learners","work_id":"63429aa6-0c8a-41f5-ac95-3960dadf41a5","ref_index":3,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2022,"title":"A literature survey of recent advances in chatbots","work_id":"a70b48fd-0211-42f2-881e-b610c0ba0e31","ref_index":4,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2021,"title":"A mathematical framework for transformer circuits","work_id":"29a1e678-f416-4561-b3dd-8b577be8fbd2","ref_index":5,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2021,"title":"Causal abstractions of neural networks","work_id":"39190767-1184-4c7f-8e56-02bea9191c0f","ref_index":6,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2022,"title":"X-Risk Analysis for AI Research","work_id":"bf80b5c8-e3d0-4a74-98cc-d60d4b26fa3d","ref_index":8,"cited_arxiv_id":"","is_internal_anchor":false}],"resolved_work":66,"snapshot_sha256":"83c8ab88bd640438564a42f60fce18d3600ee29bc819c303ce0af2f742cadc48","internal_anchors":5},"formal_canon":{"evidence_count":1,"snapshot_sha256":"9fec8a18e383908f6646ee9ad8b669c46aca0a2e59e88efff7f4105fae1dbcc8"},"author_claims":{"count":0,"strong_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"builder_version":"pith-number-builder-2026-05-17-v1"},"verdict_id":"cfd4a671-6f3d-4097-9a96-a79116a5f62e"},"signer":{"signer_id":"pith.science","signer_type":"pith_registry","key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"created_at":"2026-07-05T05:12:24Z","supersedes":[],"prev_event":null,"signature":{"signature_status":"signed_v1","algorithm":"ed25519","key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54","signature_b64":"qTWUpdWr5841CqFoaPf/CLo6SqK0Pn9xFfUqvELvw48RkqeUXH1spRCSE7JOyyIUNYJSIX87kB1A239TPKUfCQ==","signed_message":"open_graph_event_sha256_bytes","signed_at":"2026-08-10T22:16:55.493475Z"},"content_sha256":"496ee4fb962823a8323a03ef65d51b758b29de201e8515dd98dbd4c72c05c336","schema_version":"1.0","event_id":"sha256:496ee4fb962823a8323a03ef65d51b758b29de201e8515dd98dbd4c72c05c336"}],"timestamp_proofs":[],"mirror_hints":[{"mirror_type":"https","name":"Pith Resolver","base_url":"https://pith.science","bundle_url":"https://pith.science/pith/B7JM6H5UFDNTCIE5PSOVEXGQZB/bundle.json","state_url":"https://pith.science/pith/B7JM6H5UFDNTCIE5PSOVEXGQZB/state.json","well_known_bundle_url":"https://pith.science/.well-known/pith/B7JM6H5UFDNTCIE5PSOVEXGQZB/bundle.json","status":"primary"}],"public_keys":[{"key_id":"pith-v1-2026-05","algorithm":"ed25519","format":"raw","public_key_b64":"stVStoiQhXFxp4s2pdzPNoqVNBMojDU/fJ2db5S3CbM=","public_key_hex":"b2d552b68890857171a78b36a5dccf368a953413288c353f7c9d9d6f94b709b3","fingerprint_sha256_b32_first128bits":"RVFV5Z2OI2J3ZUO7ERDEBCYNKS","fingerprint_sha256_hex":"8d4b5ee74e4693bcd1df2446408b0d54","rotates_at":null,"url":"https://pith.science/pith-signing-key.json","notes":"Pith uses this Ed25519 key to sign canonical record SHA-256 digests. Verify with: ed25519_verify(public_key, message=canonical_sha256_bytes, signature=base64decode(signature_b64))."}],"merge_version":"pith-open-graph-merge-v1","built_at":"2026-08-10T22:16:55Z","links":{"resolver":"https://pith.science/pith/B7JM6H5UFDNTCIE5PSOVEXGQZB","bundle":"https://pith.science/pith/B7JM6H5UFDNTCIE5PSOVEXGQZB/bundle.json","state":"https://pith.science/pith/B7JM6H5UFDNTCIE5PSOVEXGQZB/state.json","well_known_bundle":"https://pith.science/.well-known/pith/B7JM6H5UFDNTCIE5PSOVEXGQZB/bundle.json"},"state":{"state_type":"pith_open_graph_state","state_version":"1.0","pith_number":"pith:2022:B7JM6H5UFDNTCIE5PSOVEXGQZB","merge_version":"pith-open-graph-merge-v1","event_count":2,"valid_event_count":2,"invalid_event_count":0,"equivocation_count":0,"current":{"canonical_record":{"metadata":{"abstract_canon_sha256":"63e3a75ec03b2c90703e86ed86c17d888e130ec03032fbf52ecd408fd9e5054e","cross_cats_sorted":["cs.AI","cs.CL"],"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.LG","submitted_at":"2022-11-01T17:08:44Z","title_canon_sha256":"5e96390378e8e0b98d86343f96a46f7677ac0ba7a59d5484d45b812fd5baea4b"},"schema_version":"1.0","source":{"id":"2211.00593","kind":"arxiv","version":1}},"source_aliases":[{"alias_kind":"arxiv","alias_value":"2211.00593","created_at":"2026-07-05T05:12:24Z"},{"alias_kind":"arxiv_version","alias_value":"2211.00593v1","created_at":"2026-07-05T05:12:24Z"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2211.00593","created_at":"2026-07-05T05:12:24Z"},{"alias_kind":"pith_short_12","alias_value":"B7JM6H5UFDNT","created_at":"2026-07-05T05:12:24Z"},{"alias_kind":"pith_short_16","alias_value":"B7JM6H5UFDNTCIE5","created_at":"2026-07-05T05:12:24Z"},{"alias_kind":"pith_short_8","alias_value":"B7JM6H5U","created_at":"2026-07-05T05:12:24Z"}],"graph_snapshots":[{"event_id":"sha256:496ee4fb962823a8323a03ef65d51b758b29de201e8515dd98dbd4c72c05c336","target":"graph","created_at":"2026-07-05T05:12:24Z","signer":{"key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54","signer_id":"pith.science","signer_type":"pith_registry"},"payload":{"graph_snapshot":{"author_claims":{"count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57","strong_count":0},"builder_version":"pith-number-builder-2026-05-17-v1","claims":{"count":4,"items":[{"attestation":"unclaimed","claim_id":"C1","kind":"strongest_claim","source":"verdict.strongest_claim","status":"machine_extracted","text":"Our explanation encompasses 26 attention heads grouped into 7 main classes, which we discovered using a combination of interpretability approaches relying on causal interventions. To our knowledge, this investigation is the largest end-to-end attempt at reverse-engineering a natural behavior 'in the wild' in a language model."},{"attestation":"unclaimed","claim_id":"C2","kind":"weakest_assumption","source":"verdict.weakest_assumption","status":"machine_extracted","text":"That the three quantitative criteria (faithfulness, completeness, minimality) are sufficient to certify that the identified circuit is the complete and minimal explanation rather than one of several possible circuits that achieve similar performance."},{"attestation":"unclaimed","claim_id":"C3","kind":"one_line_summary","source":"verdict.one_line_summary","status":"machine_extracted","text":"GPT-2 small solves indirect object identification via a circuit of 26 attention heads organized into seven functional classes discovered through causal interventions."},{"attestation":"unclaimed","claim_id":"C4","kind":"headline","source":"verdict.pith_extraction.headline","status":"machine_extracted","text":"GPT-2 small solves indirect object identification using a circuit of 26 attention heads in seven classes."}],"snapshot_sha256":"83622a4e2142242358d616bb24d09074ac2054bf96f77e81a564b1cd170bbc9f"},"formal_canon":{"evidence_count":1,"snapshot_sha256":"9fec8a18e383908f6646ee9ad8b669c46aca0a2e59e88efff7f4105fae1dbcc8"},"integrity":{"available":true,"clean":true,"detectors_run":[],"endpoint":"/pith/2211.00593/integrity.json","findings":[],"snapshot_sha256":"c28c3603d3b5d939e8dc4c7e95fa8dfce3d595e45f758748cecf8e644a296938","summary":{"advisory":0,"by_detector":{},"critical":0,"informational":0}},"paper":{"abstract_excerpt":"Research in mechanistic interpretability seeks to explain behaviors of machine learning models in terms of their internal components. However, most previous work either focuses on simple behaviors in small models, or describes complicated behaviors in larger models with broad strokes. In this work, we bridge this gap by presenting an explanation for how GPT-2 small performs a natural language task called indirect object identification (IOI). Our explanation encompasses 26 attention heads grouped into 7 main classes, which we discovered using a combination of interpretability approaches relying","authors_text":"Alexandre Variengien, Arthur Conmy, Buck Shlegeris, Jacob Steinhardt, Kevin Wang","cross_cats":["cs.AI","cs.CL"],"headline":"GPT-2 small solves indirect object identification using a circuit of 26 attention heads in seven classes.","license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.LG","submitted_at":"2022-11-01T17:08:44Z","title":"Interpretability in the Wild: a Circuit for Indirect Object Identification in GPT-2 small"},"references":{"count":66,"internal_anchors":5,"resolved_work":66,"sample":[{"cited_arxiv_id":"","doi":"","is_internal_anchor":false,"ref_index":3,"title":"Language models are few-shot learners","work_id":"63429aa6-0c8a-41f5-ac95-3960dadf41a5","year":1901},{"cited_arxiv_id":"","doi":"","is_internal_anchor":false,"ref_index":4,"title":"A literature survey of recent advances in chatbots","work_id":"a70b48fd-0211-42f2-881e-b610c0ba0e31","year":2022},{"cited_arxiv_id":"","doi":"","is_internal_anchor":false,"ref_index":5,"title":"A mathematical framework for transformer circuits","work_id":"29a1e678-f416-4561-b3dd-8b577be8fbd2","year":2021},{"cited_arxiv_id":"","doi":"","is_internal_anchor":false,"ref_index":6,"title":"Causal abstractions of neural networks","work_id":"39190767-1184-4c7f-8e56-02bea9191c0f","year":2021},{"cited_arxiv_id":"","doi":"","is_internal_anchor":false,"ref_index":8,"title":"X-Risk Analysis for AI Research","work_id":"bf80b5c8-e3d0-4a74-98cc-d60d4b26fa3d","year":2022}],"snapshot_sha256":"83c8ab88bd640438564a42f60fce18d3600ee29bc819c303ce0af2f742cadc48"},"source":{"id":"2211.00593","kind":"arxiv","version":1},"verdict":{"created_at":"2026-05-13T17:08:49.885716Z","id":"cfd4a671-6f3d-4097-9a96-a79116a5f62e","model_set":{"reader":"grok-4.3"},"one_line_summary":"GPT-2 small solves indirect object identification via a circuit of 26 attention heads organized into seven functional classes discovered through causal interventions.","pipeline_version":"pith-pipeline@v0.9.0","pith_extraction_headline":"GPT-2 small solves indirect object identification using a circuit of 26 attention heads in seven classes.","strongest_claim":"Our explanation encompasses 26 attention heads grouped into 7 main classes, which we discovered using a combination of interpretability approaches relying on causal interventions. To our knowledge, this investigation is the largest end-to-end attempt at reverse-engineering a natural behavior 'in the wild' in a language model.","weakest_assumption":"That the three quantitative criteria (faithfulness, completeness, minimality) are sufficient to certify that the identified circuit is the complete and minimal explanation rather than one of several possible circuits that achieve similar performance."}},"verdict_id":"cfd4a671-6f3d-4097-9a96-a79116a5f62e"}}],"author_attestations":[],"timestamp_anchors":[],"storage_attestations":[],"citation_signatures":[],"replication_records":[],"corrections":[],"mirror_hints":[],"record_created":{"event_id":"sha256:51ac6fa85937f515f21a435aba7e6d59b0cfc2ece93fbfbc71695832cd6d624c","target":"record","created_at":"2026-07-05T05:12:24Z","signer":{"key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54","signer_id":"pith.science","signer_type":"pith_registry"},"payload":{"attestation_state":"computed","canonical_record":{"metadata":{"abstract_canon_sha256":"63e3a75ec03b2c90703e86ed86c17d888e130ec03032fbf52ecd408fd9e5054e","cross_cats_sorted":["cs.AI","cs.CL"],"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.LG","submitted_at":"2022-11-01T17:08:44Z","title_canon_sha256":"5e96390378e8e0b98d86343f96a46f7677ac0ba7a59d5484d45b812fd5baea4b"},"schema_version":"1.0","source":{"id":"2211.00593","kind":"arxiv","version":1}},"canonical_sha256":"0fd2cf1fb428db31209d7c9d525cd0c870d4375341962ea0637c80732f2b4942","receipt":{"algorithm":"ed25519","builder_version":"pith-number-builder-2026-05-17-v1","canonical_sha256":"0fd2cf1fb428db31209d7c9d525cd0c870d4375341962ea0637c80732f2b4942","first_computed_at":"2026-07-05T05:12:24.537581Z","key_id":"pith-v1-2026-05","kind":"pith_receipt","last_reissued_at":"2026-07-05T05:12:24.537581Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54","receipt_version":"0.3","signature_b64":"weo05JfLya31WMljNrTtqDDjCAz5j28ntQ3WBB0lKe1debLitHzDWl+qxYMOcH8/04uNrbvHpnM59TeH53U6Cg==","signature_status":"signed_v1","signed_at":"2026-07-05T05:12:24.537999Z","signed_message":"canonical_sha256_bytes"},"source_id":"2211.00593","source_kind":"arxiv","source_version":1}}},"equivocations":[],"invalid_events":[],"applied_event_ids":["sha256:51ac6fa85937f515f21a435aba7e6d59b0cfc2ece93fbfbc71695832cd6d624c","sha256:496ee4fb962823a8323a03ef65d51b758b29de201e8515dd98dbd4c72c05c336"],"state_sha256":"c6689e2240982bc081718ca59472a7afc0c21655c219e08d6373ca94c3cc3f92"},"bundle_signature":{"signature_status":"signed_v1","algorithm":"ed25519","key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54","signature_b64":"mf2QSU36/9MmcVK6KKTHJViNCHhUAqRsSIeOXu3h89REsVflPomzzy+Ag6wXin2qPi9eYXfyiQOcEjup5V1lCA==","signed_message":"bundle_sha256_bytes","signed_at":"2026-08-10T22:16:55.500659Z","bundle_sha256":"6bfffbf1a771116f86f2e8b190bf1781ab05928c4a29470dc3309dcab6e68ef8"}}