{"paper":{"title":"Tracing Persona Vectors Through LLM Pretraining","license":"http://creativecommons.org/licenses/by-sa/4.0/","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.","cross_cats":["cs.AI"],"primary_cat":"cs.CL","authors_text":"Dominik Glandorf, Jorge Medina Moreira, Robert West, Tanja K\\\"aser, Viktor Moskvoretskii","submitted_at":"2026-05-13T10:44:23Z","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"},"claims":{"count":4,"items":[{"kind":"strongest_claim","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.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","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.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","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.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","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.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"f4ce080edf9fb110285126ed1ef2f64df392af39e3ed6806f9e718ce295fbb28"},"source":{"id":"2605.13329","kind":"arxiv","version":1},"verdict":{"id":"cf15b19f-edf8-4b04-bf62-a47520461a31","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-14T20:25:37.504219Z","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.","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","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.","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."},"references":{"count":43,"sample":[{"doi":"","year":2026,"title":"System card: Claude Mythos preview","work_id":"be586d88-b3b3-442f-af51-60367c8eed9f","ref_index":1,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2026,"title":"GPT-5.5 system card","work_id":"242862dd-044e-4cff-af24-f6897a796321","ref_index":2,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2026,"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","ref_index":3,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2025,"title":"Open Problems in Mechanistic Interpretability","work_id":"f55f2189-55b1-4a1c-acfb-a5fa7bfa9e86","ref_index":4,"cited_arxiv_id":"2501.16496","is_internal_anchor":true},{"doi":"","year":null,"title":"A longlist of theories of impact for interpretability","work_id":"ea6fb69b-11a0-4740-91f7-74959110c476","ref_index":5,"cited_arxiv_id":"","is_internal_anchor":false}],"resolved_work":43,"snapshot_sha256":"75dada02c45812900405a106ab70ba643654e8acb0c50dba61640b8ee7d7a50e","internal_anchors":5},"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"}