{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2025:UWRLGJ4UBH4HNPKLAC2CTO63RV","short_pith_number":"pith:UWRLGJ4U","schema_version":"1.0","canonical_sha256":"a5a2b3279409f876bd4b00b429bbdb8d55d4c15f258974ee05e569bbaa346744","source":{"kind":"arxiv","id":"2507.08218","version":2},"attestation_state":"computed","paper":{"title":"Simple Mechanistic Explanations for Out-Of-Context Reasoning","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.LG"],"primary_cat":"cs.CL","authors_text":"Atticus Wang, Joshua Engels, Neel Nanda, Oliver Clive-Griffin, Senthooran Rajamanoharan","submitted_at":"2025-07-10T23:47:05Z","abstract_excerpt":"Out-of-context reasoning (OOCR) is a phenomenon in which fine-tuned LLMs exhibit surprisingly deep out-of-distribution generalization. Rather than learning shallow heuristics, they implicitly internalize and act on the consequences of observations scattered throughout the fine-tuning data. In this work, we investigate this phenomenon mechanistically and find that many instances of OOCR in the literature have a simple explanation: the LoRA fine-tuning essentially adds a constant steering vector, steering the model towards a general concept. This improves performance on the fine-tuning task and "},"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":"2507.08218","kind":"arxiv","version":2},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.CL","submitted_at":"2025-07-10T23:47:05Z","cross_cats_sorted":["cs.LG"],"title_canon_sha256":"c5b46eedeef588a58a8f6302356033967a478368ac22e1b63d2d468204e29189","abstract_canon_sha256":"561b7a93dcf5f1beb3f49157aa107c26ab6edc5c2d61eac45aa6436579d79ab2"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T11:38:09.116072Z","signature_b64":"2vST0d2RTQyjUGl3lgPsPSLE12dZl9XLkh9VjfZuaWCmTi96ZYhzM3obp0voOfZ/jksShG0ygtmzvHsJe2D3BA==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"a5a2b3279409f876bd4b00b429bbdb8d55d4c15f258974ee05e569bbaa346744","last_reissued_at":"2026-07-05T11:38:09.115517Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T11:38:09.115517Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Simple Mechanistic Explanations for Out-Of-Context Reasoning","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.LG"],"primary_cat":"cs.CL","authors_text":"Atticus Wang, Joshua Engels, Neel Nanda, Oliver Clive-Griffin, Senthooran Rajamanoharan","submitted_at":"2025-07-10T23:47:05Z","abstract_excerpt":"Out-of-context reasoning (OOCR) is a phenomenon in which fine-tuned LLMs exhibit surprisingly deep out-of-distribution generalization. Rather than learning shallow heuristics, they implicitly internalize and act on the consequences of observations scattered throughout the fine-tuning data. In this work, we investigate this phenomenon mechanistically and find that many instances of OOCR in the literature have a simple explanation: the LoRA fine-tuning essentially adds a constant steering vector, steering the model towards a general concept. This improves performance on the fine-tuning task and "},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2507.08218","kind":"arxiv","version":2},"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/2507.08218/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":"2507.08218","created_at":"2026-07-05T11:38:09.115611+00:00"},{"alias_kind":"arxiv_version","alias_value":"2507.08218v2","created_at":"2026-07-05T11:38:09.115611+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2507.08218","created_at":"2026-07-05T11:38:09.115611+00:00"},{"alias_kind":"pith_short_12","alias_value":"UWRLGJ4UBH4H","created_at":"2026-07-05T11:38:09.115611+00:00"},{"alias_kind":"pith_short_16","alias_value":"UWRLGJ4UBH4HNPKL","created_at":"2026-07-05T11:38:09.115611+00:00"},{"alias_kind":"pith_short_8","alias_value":"UWRLGJ4U","created_at":"2026-07-05T11:38:09.115611+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":0,"internal_anchor_count":0,"sample":[]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/UWRLGJ4UBH4HNPKLAC2CTO63RV","json":"https://pith.science/pith/UWRLGJ4UBH4HNPKLAC2CTO63RV.json","graph_json":"https://pith.science/api/pith-number/UWRLGJ4UBH4HNPKLAC2CTO63RV/graph.json","events_json":"https://pith.science/api/pith-number/UWRLGJ4UBH4HNPKLAC2CTO63RV/events.json","paper":"https://pith.science/paper/UWRLGJ4U"},"agent_actions":{"view_html":"https://pith.science/pith/UWRLGJ4UBH4HNPKLAC2CTO63RV","download_json":"https://pith.science/pith/UWRLGJ4UBH4HNPKLAC2CTO63RV.json","view_paper":"https://pith.science/paper/UWRLGJ4U","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2507.08218&json=true","fetch_graph":"https://pith.science/api/pith-number/UWRLGJ4UBH4HNPKLAC2CTO63RV/graph.json","fetch_events":"https://pith.science/api/pith-number/UWRLGJ4UBH4HNPKLAC2CTO63RV/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/UWRLGJ4UBH4HNPKLAC2CTO63RV/action/timestamp_anchor","attest_storage":"https://pith.science/pith/UWRLGJ4UBH4HNPKLAC2CTO63RV/action/storage_attestation","attest_author":"https://pith.science/pith/UWRLGJ4UBH4HNPKLAC2CTO63RV/action/author_attestation","sign_citation":"https://pith.science/pith/UWRLGJ4UBH4HNPKLAC2CTO63RV/action/citation_signature","submit_replication":"https://pith.science/pith/UWRLGJ4UBH4HNPKLAC2CTO63RV/action/replication_record"}},"created_at":"2026-07-05T11:38:09.115611+00:00","updated_at":"2026-07-05T11:38:09.115611+00:00"}