{"state_type":"pith_open_graph_state","state_version":"1.0","pith_number":"pith:2026:5FK7HQTNIH2TDBCPZTCZTHNHAH","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":"c7f411b96e19481262bd472de5e00f7cd675e062a5443b4bf5b4fd7f7272b524","cross_cats_sorted":["cs.AI"],"license":"http://creativecommons.org/licenses/by-sa/4.0/","primary_cat":"cs.CL","submitted_at":"2026-05-13T10:44:23Z","title_canon_sha256":"553b4349d96b595ca66f5501ab112f96fb59c45acf7fcdaa8a02da0a430473a7"},"schema_version":"1.0","source":{"id":"2605.13329","kind":"arxiv","version":1}},"source_aliases":[{"alias_kind":"arxiv","alias_value":"2605.13329","created_at":"2026-05-18T02:44:48Z"},{"alias_kind":"arxiv_version","alias_value":"2605.13329v1","created_at":"2026-05-18T02:44:48Z"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2605.13329","created_at":"2026-05-18T02:44:48Z"},{"alias_kind":"pith_short_12","alias_value":"5FK7HQTNIH2T","created_at":"2026-05-18T12:33:37Z"},{"alias_kind":"pith_short_16","alias_value":"5FK7HQTNIH2TDBCP","created_at":"2026-05-18T12:33:37Z"},{"alias_kind":"pith_short_8","alias_value":"5FK7HQTN","created_at":"2026-05-18T12:33:37Z"}],"graph_snapshots":[{"event_id":"sha256:4f3967fea077a6fc29de904c95b4fbbf02b46314b2984de8326049384b01a797","target":"graph","created_at":"2026-05-18T02:44:48Z","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":"persona vectors form remarkably early -- within 0.22% of OLMo-3 pretraining -- and remain effective for steering the fully post-trained instruct models. Although core representations are formed early on, persona vectors continue to refine geometrically and semantically throughout pretraining."},{"attestation":"unclaimed","claim_id":"C2","kind":"weakest_assumption","source":"verdict.weakest_assumption","status":"machine_extracted","text":"That the linear directions identified at early checkpoints represent the same high-level personas as those in the final model and that the elicitation methods isolate these without being confounded by other training dynamics."},{"attestation":"unclaimed","claim_id":"C3","kind":"one_line_summary","source":"verdict.one_line_summary","status":"machine_extracted","text":"Persona vectors form within the first 0.22% of LLM pretraining and remain effective for steering post-trained models, with continued refinement and transfer to other models."},{"attestation":"unclaimed","claim_id":"C4","kind":"headline","source":"verdict.pith_extraction.headline","status":"machine_extracted","text":"Persona vectors for traits like sycophancy emerge within the first 0.22 percent of LLM pretraining and remain usable for steering the final model."}],"snapshot_sha256":"f4ce080edf9fb110285126ed1ef2f64df392af39e3ed6806f9e718ce295fbb28"},"formal_canon":{"evidence_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"paper":{"abstract_excerpt":"How large language models internally represent high-level behaviors is a core interpretability question with direct relevance to AI safety: it determines what we can detect, audit, or intervene on. Recent work has shown that traits such as evil or sycophancy correspond to linear directions in the internal activations, the so-called persona vectors. Although these vectors are now routinely utilized to inspect and steer model behavior in safety-relevant settings, how these representations are formed during training remains unknown. To address this gap, we trace persona vectors across the pretrai","authors_text":"Dominik Glandorf, Jorge Medina Moreira, Robert West, Tanja K\\\"aser, Viktor Moskvoretskii","cross_cats":["cs.AI"],"headline":"Persona vectors for traits like sycophancy emerge within the first 0.22 percent of LLM pretraining and remain usable for steering the final model.","license":"http://creativecommons.org/licenses/by-sa/4.0/","primary_cat":"cs.CL","submitted_at":"2026-05-13T10:44:23Z","title":"Tracing Persona Vectors Through LLM Pretraining"},"references":{"count":43,"internal_anchors":5,"resolved_work":43,"sample":[{"cited_arxiv_id":"","doi":"","is_internal_anchor":false,"ref_index":1,"title":"System card: Claude Mythos preview","work_id":"be586d88-b3b3-442f-af51-60367c8eed9f","year":2026},{"cited_arxiv_id":"","doi":"","is_internal_anchor":false,"ref_index":2,"title":"GPT-5.5 system card","work_id":"242862dd-044e-4cff-af24-f6897a796321","year":2026},{"cited_arxiv_id":"","doi":"","is_internal_anchor":false,"ref_index":3,"title":"Jan Betley, Daniel Chee Hian Tan, Niels Warncke, Anna Sztyber-Betley, Xuchan Bao, Mart´ın Soto, Nathan Labenz, and Owain Evans","work_id":"c2d55f47-a2d5-4219-a7d2-094e9a8839f1","year":2026},{"cited_arxiv_id":"2501.16496","doi":"","is_internal_anchor":true,"ref_index":4,"title":"Open Problems in Mechanistic Interpretability","work_id":"f55f2189-55b1-4a1c-acfb-a5fa7bfa9e86","year":2025},{"cited_arxiv_id":"","doi":"","is_internal_anchor":false,"ref_index":5,"title":"A longlist of theories of impact for interpretability","work_id":"ea6fb69b-11a0-4740-91f7-74959110c476","year":null}],"snapshot_sha256":"75dada02c45812900405a106ab70ba643654e8acb0c50dba61640b8ee7d7a50e"},"source":{"id":"2605.13329","kind":"arxiv","version":1},"verdict":{"created_at":"2026-05-14T20:25:37.504219Z","id":"cf15b19f-edf8-4b04-bf62-a47520461a31","model_set":{"reader":"grok-4.3"},"one_line_summary":"Persona vectors form within the first 0.22% of LLM pretraining and remain effective for steering post-trained models, with continued refinement and transfer to other models.","pipeline_version":"pith-pipeline@v0.9.0","pith_extraction_headline":"Persona vectors for traits like sycophancy emerge within the first 0.22 percent of LLM pretraining and remain usable for steering the final model.","strongest_claim":"persona vectors form remarkably early -- within 0.22% of OLMo-3 pretraining -- and remain effective for steering the fully post-trained instruct models. Although core representations are formed early on, persona vectors continue to refine geometrically and semantically throughout pretraining.","weakest_assumption":"That the linear directions identified at early checkpoints represent the same high-level personas as those in the final model and that the elicitation methods isolate these without being confounded by other training dynamics."}},"verdict_id":"cf15b19f-edf8-4b04-bf62-a47520461a31"}}],"author_attestations":[],"timestamp_anchors":[],"storage_attestations":[],"citation_signatures":[],"replication_records":[],"corrections":[],"mirror_hints":[],"record_created":{"event_id":"sha256:2fee8a0a3747a9857e9d8783a893fad82ecdae8eae74052511eac26f00026f02","target":"record","created_at":"2026-05-18T02:44:48Z","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":"c7f411b96e19481262bd472de5e00f7cd675e062a5443b4bf5b4fd7f7272b524","cross_cats_sorted":["cs.AI"],"license":"http://creativecommons.org/licenses/by-sa/4.0/","primary_cat":"cs.CL","submitted_at":"2026-05-13T10:44:23Z","title_canon_sha256":"553b4349d96b595ca66f5501ab112f96fb59c45acf7fcdaa8a02da0a430473a7"},"schema_version":"1.0","source":{"id":"2605.13329","kind":"arxiv","version":1}},"canonical_sha256":"e955f3c26d41f531844fccc5999da701cfadd98e7ef96b2818fb79257165c9fc","receipt":{"algorithm":"ed25519","builder_version":"pith-number-builder-2026-05-17-v1","canonical_sha256":"e955f3c26d41f531844fccc5999da701cfadd98e7ef96b2818fb79257165c9fc","first_computed_at":"2026-05-18T02:44:48.576627Z","key_id":"pith-v1-2026-05","kind":"pith_receipt","last_reissued_at":"2026-05-18T02:44:48.576627Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54","receipt_version":"0.3","signature_b64":"CRuBfOfynbSwxZUyF9BiLqjQsAkWSHPYdRxiEV2fYd5gHQ0bHNp9Zgwp8jwKmhtRldV3BHLL99tpGg2RrZ1hDA==","signature_status":"signed_v1","signed_at":"2026-05-18T02:44:48.577063Z","signed_message":"canonical_sha256_bytes"},"source_id":"2605.13329","source_kind":"arxiv","source_version":1}}},"equivocations":[],"invalid_events":[],"applied_event_ids":["sha256:2fee8a0a3747a9857e9d8783a893fad82ecdae8eae74052511eac26f00026f02","sha256:4f3967fea077a6fc29de904c95b4fbbf02b46314b2984de8326049384b01a797"],"state_sha256":"ae5059906a297a9950bf9bcd8c8c142a52968800b6b794c55a0df1b31ec10efa"}