{"bundle_type":"pith_open_graph_bundle","bundle_version":"1.0","pith_number":"pith:2023:4A6VNY3DWS2C6DZRC7XC7LQPBU","short_pith_number":"pith:4A6VNY3D","canonical_record":{"source":{"id":"2304.05969","kind":"arxiv","version":2},"metadata":{"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.LG","submitted_at":"2023-04-12T16:46:43Z","cross_cats_sorted":[],"title_canon_sha256":"bda85cc9f605bea8fb0a871b936d9297d3f21040aca6ee83a2b6eb3ab2837a8c","abstract_canon_sha256":"f1f8ee0e12cf4262f1dae11d3bbb394d0ce2b8a5a24eaec7fe27148b6c8d664f"},"schema_version":"1.0"},"canonical_sha256":"e03d56e363b4b42f0f3117ee2fae0f0d34b92b58ea6fd29b76246a68523d9b51","source":{"kind":"arxiv","id":"2304.05969","version":2},"source_aliases":[{"alias_kind":"arxiv","alias_value":"2304.05969","created_at":"2026-05-17T23:38:46Z"},{"alias_kind":"arxiv_version","alias_value":"2304.05969v2","created_at":"2026-05-17T23:38:46Z"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2304.05969","created_at":"2026-05-17T23:38:46Z"},{"alias_kind":"pith_short_12","alias_value":"4A6VNY3DWS2C","created_at":"2026-05-18T12:33:33Z"},{"alias_kind":"pith_short_16","alias_value":"4A6VNY3DWS2C6DZR","created_at":"2026-05-18T12:33:33Z"},{"alias_kind":"pith_short_8","alias_value":"4A6VNY3D","created_at":"2026-05-18T12:33:33Z"}],"events":[{"event_type":"record_created","subject_pith_number":"pith:2023:4A6VNY3DWS2C6DZRC7XC7LQPBU","target":"record","payload":{"canonical_record":{"source":{"id":"2304.05969","kind":"arxiv","version":2},"metadata":{"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.LG","submitted_at":"2023-04-12T16:46:43Z","cross_cats_sorted":[],"title_canon_sha256":"bda85cc9f605bea8fb0a871b936d9297d3f21040aca6ee83a2b6eb3ab2837a8c","abstract_canon_sha256":"f1f8ee0e12cf4262f1dae11d3bbb394d0ce2b8a5a24eaec7fe27148b6c8d664f"},"schema_version":"1.0"},"canonical_sha256":"e03d56e363b4b42f0f3117ee2fae0f0d34b92b58ea6fd29b76246a68523d9b51","receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-05-17T23:38:46.815123Z","signature_b64":"Bo78BkWhxOlnD8ESzLEpTOb/UTmNEsM7lsj8O/8GbARCw2OutwxdKIbiqNawPKNGJugbd6iog3pTteUDurtcCg==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"e03d56e363b4b42f0f3117ee2fae0f0d34b92b58ea6fd29b76246a68523d9b51","last_reissued_at":"2026-05-17T23:38:46.814622Z","signature_status":"signed_v1","first_computed_at":"2026-05-17T23:38:46.814622Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"source_kind":"arxiv","source_id":"2304.05969","source_version":2,"attestation_state":"computed"},"signer":{"signer_id":"pith.science","signer_type":"pith_registry","key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"created_at":"2026-05-17T23:38:46Z","supersedes":[],"prev_event":null,"signature":{"signature_status":"signed_v1","algorithm":"ed25519","key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54","signature_b64":"JOdg5ucSoBMXGrS5QMm2yJRP7PGKKW7/eu5SzQB+EmQ17HroiODF7Z5lMWMHmT2XJ5SE4vXRb9lPp6uVzwCpBQ==","signed_message":"open_graph_event_sha256_bytes","signed_at":"2026-05-30T07:45:15.808153Z"},"content_sha256":"aa958871057d438efb26b69e3ce7ee870ada231f7f80ab83f9c85c6c2ab1f605","schema_version":"1.0","event_id":"sha256:aa958871057d438efb26b69e3ce7ee870ada231f7f80ab83f9c85c6c2ab1f605"},{"event_type":"graph_snapshot","subject_pith_number":"pith:2023:4A6VNY3DWS2C6DZRC7XC7LQPBU","target":"graph","payload":{"graph_snapshot":{"paper":{"title":"Localizing Model Behavior with Path Patching","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"Path patching lets researchers test whether a neural network's behavior is localized to a specific set of paths through its components.","cross_cats":[],"primary_cat":"cs.LG","authors_text":"Aryaman Arora, Chris MacLeod, Lucas Sato, Nicholas Goldowsky-Dill","submitted_at":"2023-04-12T16:46:43Z","abstract_excerpt":"Localizing behaviors of neural networks to a subset of the network's components or a subset of interactions between components is a natural first step towards analyzing network mechanisms and possible failure modes. Existing work is often qualitative and ad-hoc, and there is no consensus on the appropriate way to evaluate localization claims. We introduce path patching, a technique for expressing and quantitatively testing a natural class of hypotheses expressing that behaviors are localized to a set of paths. We refine an explanation of induction heads, characterize a behavior of GPT-2, and o"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"We introduce path patching, a technique for expressing and quantitatively testing a natural class of hypotheses expressing that behaviors are localized to a set of paths.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"That altering activations along hypothesized paths isolates the causal contribution of those paths without introducing significant side effects or confounding interactions elsewhere in the network.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"Path patching provides a method to express and quantitatively test hypotheses that neural network behaviors are localized to sets of paths.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"Path patching lets researchers test whether a neural network's behavior is localized to a specific set of paths through its components.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"c2489ab87e49d7fccc8da9d72dd49ca1e04cffe42068742d3c122d4ae90a7a79"},"source":{"id":"2304.05969","kind":"arxiv","version":2},"verdict":{"id":"100ade03-f888-4117-a71b-b51d9fd5b9f3","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-16T19:34:59.863980Z","strongest_claim":"We introduce path patching, a technique for expressing and quantitatively testing a natural class of hypotheses expressing that behaviors are localized to a set of paths.","one_line_summary":"Path patching provides a method to express and quantitatively test hypotheses that neural network behaviors are localized to sets of paths.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"That altering activations along hypothesized paths isolates the causal contribution of those paths without introducing significant side effects or confounding interactions elsewhere in the network.","pith_extraction_headline":"Path patching lets researchers test whether a neural network's behavior is localized to a specific set of paths through its components."},"references":{"count":48,"sample":[{"doi":"","year":2022,"title":"Interpretability in the Wild: a Circuit for Indirect Object Identification in GPT-2 small , author=. 2022 , eprint=","work_id":"d1c74c4a-b873-4046-bc5d-926934ef2064","ref_index":1,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2023,"title":"Towards Automated Circuit Discovery for Mechanistic Interpretability , author=. 2023 , archivePrefix=","work_id":"a94d79c6-5d9b-4836-ada5-b9b99a0859e7","ref_index":2,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":null,"title":"In-context Learning and Induction Heads , author=. ArXiv , year=","work_id":"1168a9ce-2ab3-4e49-a7de-e864432aacec","ref_index":3,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":null,"title":"Zoom in: An introduction to circuits , author=. Distill , volume=","work_id":"2c5f7a19-7432-4615-8303-f2fc83a88db7","ref_index":4,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2023,"title":"A Toy Model of Universality: Reverse Engineering How Networks Learn Group Operations , author=. 2023 , eprint=","work_id":"60971e8c-14f6-4fb0-8249-0877c17ead9d","ref_index":5,"cited_arxiv_id":"","is_internal_anchor":false}],"resolved_work":48,"snapshot_sha256":"351411bbbec4beef421b8ed3448e5fe09fc93d5181400afd7f73e70c5dcd8a2a","internal_anchors":2},"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"},"verdict_id":"100ade03-f888-4117-a71b-b51d9fd5b9f3"},"signer":{"signer_id":"pith.science","signer_type":"pith_registry","key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"created_at":"2026-05-17T23:38:46Z","supersedes":[],"prev_event":null,"signature":{"signature_status":"signed_v1","algorithm":"ed25519","key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54","signature_b64":"aMsocOpvDmHhoZ9vCdwjwHpihDg+O5KItFDgD/ItIbMEg25gBM9pGOYdmzxIZrkr81a77TnmO7crIPOaO2o5DA==","signed_message":"open_graph_event_sha256_bytes","signed_at":"2026-05-30T07:45:15.808711Z"},"content_sha256":"6a345be4d19b90ab80440f156964f741b7748acd8d041ab4c6358a9f5cd0ca18","schema_version":"1.0","event_id":"sha256:6a345be4d19b90ab80440f156964f741b7748acd8d041ab4c6358a9f5cd0ca18"}],"timestamp_proofs":[],"mirror_hints":[{"mirror_type":"https","name":"Pith Resolver","base_url":"https://pith.science","bundle_url":"https://pith.science/pith/4A6VNY3DWS2C6DZRC7XC7LQPBU/bundle.json","state_url":"https://pith.science/pith/4A6VNY3DWS2C6DZRC7XC7LQPBU/state.json","well_known_bundle_url":"https://pith.science/.well-known/pith/4A6VNY3DWS2C6DZRC7XC7LQPBU/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":"2026-05-30T07:45:15Z","links":{"resolver":"https://pith.science/pith/4A6VNY3DWS2C6DZRC7XC7LQPBU","bundle":"https://pith.science/pith/4A6VNY3DWS2C6DZRC7XC7LQPBU/bundle.json","state":"https://pith.science/pith/4A6VNY3DWS2C6DZRC7XC7LQPBU/state.json","well_known_bundle":"https://pith.science/.well-known/pith/4A6VNY3DWS2C6DZRC7XC7LQPBU/bundle.json"},"state":{"state_type":"pith_open_graph_state","state_version":"1.0","pith_number":"pith:2023:4A6VNY3DWS2C6DZRC7XC7LQPBU","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":"f1f8ee0e12cf4262f1dae11d3bbb394d0ce2b8a5a24eaec7fe27148b6c8d664f","cross_cats_sorted":[],"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.LG","submitted_at":"2023-04-12T16:46:43Z","title_canon_sha256":"bda85cc9f605bea8fb0a871b936d9297d3f21040aca6ee83a2b6eb3ab2837a8c"},"schema_version":"1.0","source":{"id":"2304.05969","kind":"arxiv","version":2}},"source_aliases":[{"alias_kind":"arxiv","alias_value":"2304.05969","created_at":"2026-05-17T23:38:46Z"},{"alias_kind":"arxiv_version","alias_value":"2304.05969v2","created_at":"2026-05-17T23:38:46Z"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2304.05969","created_at":"2026-05-17T23:38:46Z"},{"alias_kind":"pith_short_12","alias_value":"4A6VNY3DWS2C","created_at":"2026-05-18T12:33:33Z"},{"alias_kind":"pith_short_16","alias_value":"4A6VNY3DWS2C6DZR","created_at":"2026-05-18T12:33:33Z"},{"alias_kind":"pith_short_8","alias_value":"4A6VNY3D","created_at":"2026-05-18T12:33:33Z"}],"graph_snapshots":[{"event_id":"sha256:6a345be4d19b90ab80440f156964f741b7748acd8d041ab4c6358a9f5cd0ca18","target":"graph","created_at":"2026-05-17T23:38:46Z","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":"We introduce path patching, a technique for expressing and quantitatively testing a natural class of hypotheses expressing that behaviors are localized to a set of paths."},{"attestation":"unclaimed","claim_id":"C2","kind":"weakest_assumption","source":"verdict.weakest_assumption","status":"machine_extracted","text":"That altering activations along hypothesized paths isolates the causal contribution of those paths without introducing significant side effects or confounding interactions elsewhere in the network."},{"attestation":"unclaimed","claim_id":"C3","kind":"one_line_summary","source":"verdict.one_line_summary","status":"machine_extracted","text":"Path patching provides a method to express and quantitatively test hypotheses that neural network behaviors are localized to sets of paths."},{"attestation":"unclaimed","claim_id":"C4","kind":"headline","source":"verdict.pith_extraction.headline","status":"machine_extracted","text":"Path patching lets researchers test whether a neural network's behavior is localized to a specific set of paths through its components."}],"snapshot_sha256":"c2489ab87e49d7fccc8da9d72dd49ca1e04cffe42068742d3c122d4ae90a7a79"},"formal_canon":{"evidence_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"paper":{"abstract_excerpt":"Localizing behaviors of neural networks to a subset of the network's components or a subset of interactions between components is a natural first step towards analyzing network mechanisms and possible failure modes. Existing work is often qualitative and ad-hoc, and there is no consensus on the appropriate way to evaluate localization claims. We introduce path patching, a technique for expressing and quantitatively testing a natural class of hypotheses expressing that behaviors are localized to a set of paths. We refine an explanation of induction heads, characterize a behavior of GPT-2, and o","authors_text":"Aryaman Arora, Chris MacLeod, Lucas Sato, Nicholas Goldowsky-Dill","cross_cats":[],"headline":"Path patching lets researchers test whether a neural network's behavior is localized to a specific set of paths through its components.","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.LG","submitted_at":"2023-04-12T16:46:43Z","title":"Localizing Model Behavior with Path Patching"},"references":{"count":48,"internal_anchors":2,"resolved_work":48,"sample":[{"cited_arxiv_id":"","doi":"","is_internal_anchor":false,"ref_index":1,"title":"Interpretability in the Wild: a Circuit for Indirect Object Identification in GPT-2 small , author=. 2022 , eprint=","work_id":"d1c74c4a-b873-4046-bc5d-926934ef2064","year":2022},{"cited_arxiv_id":"","doi":"","is_internal_anchor":false,"ref_index":2,"title":"Towards Automated Circuit Discovery for Mechanistic Interpretability , author=. 2023 , archivePrefix=","work_id":"a94d79c6-5d9b-4836-ada5-b9b99a0859e7","year":2023},{"cited_arxiv_id":"","doi":"","is_internal_anchor":false,"ref_index":3,"title":"In-context Learning and Induction Heads , author=. ArXiv , year=","work_id":"1168a9ce-2ab3-4e49-a7de-e864432aacec","year":null},{"cited_arxiv_id":"","doi":"","is_internal_anchor":false,"ref_index":4,"title":"Zoom in: An introduction to circuits , author=. Distill , volume=","work_id":"2c5f7a19-7432-4615-8303-f2fc83a88db7","year":null},{"cited_arxiv_id":"","doi":"","is_internal_anchor":false,"ref_index":5,"title":"A Toy Model of Universality: Reverse Engineering How Networks Learn Group Operations , author=. 2023 , eprint=","work_id":"60971e8c-14f6-4fb0-8249-0877c17ead9d","year":2023}],"snapshot_sha256":"351411bbbec4beef421b8ed3448e5fe09fc93d5181400afd7f73e70c5dcd8a2a"},"source":{"id":"2304.05969","kind":"arxiv","version":2},"verdict":{"created_at":"2026-05-16T19:34:59.863980Z","id":"100ade03-f888-4117-a71b-b51d9fd5b9f3","model_set":{"reader":"grok-4.3"},"one_line_summary":"Path patching provides a method to express and quantitatively test hypotheses that neural network behaviors are localized to sets of paths.","pipeline_version":"pith-pipeline@v0.9.0","pith_extraction_headline":"Path patching lets researchers test whether a neural network's behavior is localized to a specific set of paths through its components.","strongest_claim":"We introduce path patching, a technique for expressing and quantitatively testing a natural class of hypotheses expressing that behaviors are localized to a set of paths.","weakest_assumption":"That altering activations along hypothesized paths isolates the causal contribution of those paths without introducing significant side effects or confounding interactions elsewhere in the network."}},"verdict_id":"100ade03-f888-4117-a71b-b51d9fd5b9f3"}}],"author_attestations":[],"timestamp_anchors":[],"storage_attestations":[],"citation_signatures":[],"replication_records":[],"corrections":[],"mirror_hints":[],"record_created":{"event_id":"sha256:aa958871057d438efb26b69e3ce7ee870ada231f7f80ab83f9c85c6c2ab1f605","target":"record","created_at":"2026-05-17T23:38:46Z","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":"f1f8ee0e12cf4262f1dae11d3bbb394d0ce2b8a5a24eaec7fe27148b6c8d664f","cross_cats_sorted":[],"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.LG","submitted_at":"2023-04-12T16:46:43Z","title_canon_sha256":"bda85cc9f605bea8fb0a871b936d9297d3f21040aca6ee83a2b6eb3ab2837a8c"},"schema_version":"1.0","source":{"id":"2304.05969","kind":"arxiv","version":2}},"canonical_sha256":"e03d56e363b4b42f0f3117ee2fae0f0d34b92b58ea6fd29b76246a68523d9b51","receipt":{"algorithm":"ed25519","builder_version":"pith-number-builder-2026-05-17-v1","canonical_sha256":"e03d56e363b4b42f0f3117ee2fae0f0d34b92b58ea6fd29b76246a68523d9b51","first_computed_at":"2026-05-17T23:38:46.814622Z","key_id":"pith-v1-2026-05","kind":"pith_receipt","last_reissued_at":"2026-05-17T23:38:46.814622Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54","receipt_version":"0.3","signature_b64":"Bo78BkWhxOlnD8ESzLEpTOb/UTmNEsM7lsj8O/8GbARCw2OutwxdKIbiqNawPKNGJugbd6iog3pTteUDurtcCg==","signature_status":"signed_v1","signed_at":"2026-05-17T23:38:46.815123Z","signed_message":"canonical_sha256_bytes"},"source_id":"2304.05969","source_kind":"arxiv","source_version":2}}},"equivocations":[],"invalid_events":[],"applied_event_ids":["sha256:aa958871057d438efb26b69e3ce7ee870ada231f7f80ab83f9c85c6c2ab1f605","sha256:6a345be4d19b90ab80440f156964f741b7748acd8d041ab4c6358a9f5cd0ca18"],"state_sha256":"016905877f47a5554dcad5191179a59a71fc50f41479366e9bcd45790308a8ef"},"bundle_signature":{"signature_status":"signed_v1","algorithm":"ed25519","key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54","signature_b64":"V9Z+L26Q4YFVrBN8w2LgxinES9HCqkrrHCAXSXcBEftastDFv2/rnokVGwxSN1NLWoq1L1XeAEappINuvQeKBw==","signed_message":"bundle_sha256_bytes","signed_at":"2026-05-30T07:45:15.811296Z","bundle_sha256":"19e25189cd73b4f60b96a64d950b1f43091304e19b773b2cdb96e4bd272a656f"}}