Summary
Moskvoretskii et al. (EPFL) trace the formation of persona vectors (linear directions in activation space corresponding to high-level behavioral traits such as "evil," sycophancy, impoliteness, and humor) across the full pretraining run of OLMo-3-7B, with replication on Apertus-8B. The central finding is that three of the four vectors (Evil, Sycophantic, Impolite) can be extracted and steer the base model from the checkpoint at 0.22% of pretraining tokens (12.6B) on OLMo-3. The paper also reports that pretraining-stage vectors remain effective for steering the post-trained Instruct model, but its per-checkpoint tables qualify this by trait (Key results). Although the core direction is present from very early checkpoints, the vectors continue to refine both geometrically (increasing alignment with the final vector) and semantically (shifting facet profiles) throughout the rest of pretraining. Different elicitation methods recover steerable vectors but emphasize distinct facets of the same persona. The authors conclude that persona representations originate in pretraining rather than in post-training.
Method
The authors use the publicly available pretraining checkpoints of OLMo-3-7B (17 checkpoints with denser sampling early in training) and replicate key results on Apertus-8B.
Persona vector extraction. For each checkpoint and target trait, they elicit contrasting positive and negative continuations from a third-person character description (adapted from the Chen et al. system-prompt setup for base models), keep only generations a GPT-4.1-mini judge scores as both on-trait and coherent, and compute a difference-of-means vector in the residual stream. The layer is fixed per trait (16 or 20), chosen from prior findings rather than tuned per checkpoint, and the steering coefficient is held fixed across checkpoints.
Steering protocol. To test a vector extracted at checkpoint τ on a target model M (which may be a later checkpoint or a post-trained variant), they add a scaled version of the vector to the residual stream during generation. Scaling is normalized relative to the target model's activation norm at that layer for comparability across training stages.
Elicitation comparison. For the Evil persona on the final checkpoint, two further discourse types (dialogue and narration) and a pooled combination are compared against the description method.
Replication. The main analyses (emergence, transfer, refinement) are repeated on Apertus-8B using 15 of its checkpoints.
Key results
Early emergence (RQ1). Evil, Sycophantic and Impolite vectors are first extractable at 12.6B tokens (0.22% of OLMo-3-7B pretraining) and steer that same checkpoint significantly (Evil +9.7, Sycophantic +12.1, Impolite +16.9 points of judge-scored trait expression; Appendix K, Table 8). The two earlier sampled checkpoints fail the coherence or trait filter. Humorous is the exception: no vector until 21B tokens, and steering effects on the base checkpoints stay under +9 through pretraining. The paper puts it as each persona having its own onset and ceiling, with humor "barely" emerging by the end of pretraining.
Transfer to the post-trained model. The paper states that vectors from early pretraining checkpoints remain largely effective on OLMo-3-7B-Instruct, with persona suppression concentrated at the DPO stage and RLVR adding little. Its per-checkpoint table (Appendix K, Table 11; Instruct coefficients retuned separately) shows the transfer is trait-dependent. Sycophantic vectors transfer from the earliest checkpoint (+4.7 at 12.6B against +7.1 for the final vector). Evil vectors transfer significantly but weakly at first (+0.1 at 12.6B, under +6 before 5.9T tokens, against +11.3 for the final vector). Impolite vectors from checkpoints before 2.97T tokens (about half of pretraining) have no significant effect on Instruct. Humorous effects stay under +2 throughout.
Continued refinement during pretraining (RQ2). Even after the core direction appears, persona vectors continue to evolve:
- Geometric refinement: cosine similarity between a checkpoint's vector and the final-pretraining vector rises steadily from about 0.3 at the earliest extractable checkpoint. Step-to-step movement is largest early.
- Semantic refinement: facet profiles for Evil and Sycophantic are largely stable, but individual facets move with the geometry (sadism grows in the Evil generations, indirect sycophancy declines).
Elicitation method effects. For Evil, all three elicitation strategies (description, dialogue, narration) produce vectors that steer significantly on their own evaluation set (Table 1). Their pairwise cosine similarities are all below 0.5, and they emphasize different roots of evil.
Replication on Apertus-8B. The qualitative pattern holds: vectors exist from the earliest available checkpoint (~1.4% of pretraining), transfer to Instruct, and refine over training. The authors note differences. Evil vectors from before 13T tokens become nearly ineffective on Apertus-Instruct, and the decline in indirect sycophancy does not replicate.
Safety implications. The authors argue that because persona representations form early and persist through alignment, pretraining itself is the natural place to intervene on them.
Why it matters
The authors present this as the first direct demonstration of when persona vectors form during training. It supports the Persona Selection Model (PSM) claim that persona representations are a pretraining phenomenon, with post-training modulating their expression. The support is trait-dependent. Only some very early vectors still steer the Instruct model appreciably, and the study covers four traits in two open models.
It introduces a new structural shape to the persona-selection cluster: pretraining temporal formation / crystallization. Prior instantiations have focused on inference-time selection, activation-level toolkits, prompt-level interventions, and fine-tuning objective effects. This finding adds a developmental timeline: most of the representations are extractable very early and then undergo prolonged refinement.
The paper itself argues for pretraining as the place to intervene.
interpretive tensions
- How early is "early"? 0.22% is where extraction first succeeds. The paper calls its emergence times lower bounds, but as onset times they are upper bounds. Extraction needs coherent persona-expressing text, and the earlier checkpoints fail that filter, so the directions may exist earlier than the pipeline can detect.
- Direction vs. facet. The earliest extractable vectors sit at cosine about 0.3 from the final direction yet still steer the base model. Direction and facets both keep refining. This complicates simple "the persona vector is fixed at step X" readings.
- Elicitation dependence. The finding that different prompts recover different facets of the "same" persona is both a methodological caution and a substantive result about the geometry of persona space.
- Scope of traits. The study focuses on four specific traits. Whether the same extreme early-formation pattern holds for other high-level behavioral directions (or for more complex, multi-faceted personas) remains open.
concepts
- Persona selection — Introduces the new structural role of pretraining temporal formation / crystallization. The finding shows that most of the studied persona directions are already extractable from base models within the first fraction of a percent of pretraining and then undergo extended geometric and semantic refinement. This is the first wiki entry to measure the developmental timeline of persona vectors directly.
The new shape (pretraining crystallization) is held at one example. It would be a natural candidate for codification under the concept once a second structurally comparable measurement of persona-vector formation timing appears.
threads
- Strengthens the mechanistic grounding for the persona-selection account of emergent misalignment, sycophancy, and related phenomena (many of which are already linked to persona vectors in the wiki).
- Raises the priority of pretraining-stage interpretability and intervention research within the broader persona-selection and safety agenda.
sources
- Moskvoretskii, V., Glandorf, D., Medina Moreira, J., Käser, T., & West, R. (2026). Tracing Persona Vectors Through LLM Pretraining. arXiv:2605.13329.
- Full primary source cached at
cache/papers/source-2026-persona-vectors-pretraining-moskvoretskii.{pdf,md}. The markdown flattens the Appendix K tables; per-checkpoint values above were read from the PDF text layer. - Related: Chen et al. (2025) "Persona Vectors" (the work this paper builds directly upon); Marks, Lindsey & Olah (2026) Persona Selection Model.