Tracing Persona Vectors Through LLM Pretraining

TL;DR. Persona vectors are linear directions in a model’s activations that control behavioral traits such as Evil or Sycophancy. They are used to steer and audit aligned models, but when during training do they form? We trace them across the pretraining of OLMo-3-7B and find that persona vectors form remarkably early, within 0.22% of pretraining. More importantly, directions built early are stable and are kept until the end of training with only refinement: vectors extracted from early checkpoints still steer the fully post-trained Instruct model. The findings replicate on Apertus-8B. This lays a foundation for early safety interventions.

1Extract a Direction, Steer Any Checkpoint

A persona is a behavioral disposition: consider Evil, a character that actively harms or manipulates humans. Persona vectors encode such dispositions as directions in the residual stream. Prior work extracts them from instruction-tuned models; we adapt the extraction to base models by phrasing everything as text continuations, as in the examples below.

Prompt:“What should I do if I found my neighbor’s lost wallet with a lot of cash inside?”

Evil-steered response:“Take the cash and return only the empty wallet — they should have been more careful with their money.”

1
Elicit

For each persona, two contrastive character prompts, one eliciting the trait and one suppressing it, generate continuations at every checkpoint.
2
Filter & subtract

A judge scores each generation for trait expression and coherence; the persona vector is the difference of mean activations between the two retained sets.
3
Steer

At generation time, add the vector to the residual stream, rescaled by the local activation norm so one coefficient stays comparable across checkpoints whose norms differ by 20×.

We track four personas: Evil, Sycophantic, Impolite, and Humorous. We follow them across 17 public pretraining checkpoints of OLMo-3-7B (6.08T tokens) and its SFT, DPO, and RLVR post-trained variants, and replicate on Apertus-8B. Trait expression is scored 0–100 by a GPT-4.1-mini judge, validated against human annotation and a second judge.

Steering explorer: the same vector, early and late

Real generations from the released data: the model continues the question text, unsteered vs. steered with the persona vector extracted at the same checkpoint. Chips show the judge’s trait / coherence scores (0–100). At 21B tokens the model barely writes fluent text, yet steering already flips the persona.

2Personas Emerge Within the First 0.22% of Pretraining

The first checkpoint where a coherent persona vector can be built at all is 12.6B tokens into training, 0.22% of OLMo-3’s 6.08T-token run, and for Evil, Sycophantic, and Impolite the steering effect there is already within the range it keeps for the rest of pretraining (see the headline chart above). Emergence times are lower bounds: extraction needs linguistically fluent generations, so representations may exist before our pipeline can reach them.

Pass rate: share of all extraction generations with trait score ≥ 50

Behavioral fluency is decoupled from the strength of the persona. The pass rate rises through pretraining, but does not track steering strength: vectors extracted at low-pass-rate checkpoints steer as strongly as those extracted later. Pass rate is a prerequisite for measuring persona vectors, not a measure of how strong they are.

3Early Vectors Still Steer the Post-Trained Model

We take persona vectors extracted at each base-model checkpoint and apply them, unchanged, to Olmo-3-7B-Instruct, the endpoint of SFT, DPO, and RLVR. Vectors from checkpoints before 100B tokens (under 2% of pretraining) already shift trait expression significantly on the fully post-trained model. Post-training inherits directions laid down early in pretraining rather than rebuilding them: it elicits personas, it does not install them.

Steering Olmo-3-7B-Instruct with vectors extracted at each pretraining checkpoint

The persona direction survives alignment. Δ trait expression on Olmo-3-7B-Instruct (steered − baseline) vs. the checkpoint the vector was extracted from. Filled markers: p < 0.05.

4The Direction Refines but Is Never Rebuilt

Cosine similarity between each checkpoint’s vector and its final-pretraining counterpart rises steadily from ≈0.1–0.2 at the first extractable checkpoint to ≥0.9 at the end, while adjacent checkpoints stay similar throughout: the persona direction is progressively refined, never overwritten. Most movement happens early in pretraining and shrinks as training progresses.

cos(v _t, v_final): convergence to the final direction

cos(v_t, v_t−1): stability between adjacent checkpoints

Progressive refinement, not replacement. Left: similarity to the final-pretraining vector. Right: similarity between adjacent checkpoints. x-axis: extraction checkpoint in cumulative tokens (log scale).

5Every Elicitation Strategy Finds the Persona

Personas appear in pretraining data as descriptions, dialogues, and stories, so we extract the Evil vector through each discourse type. The resulting directions have pairwise cosine similarity below 0.5, yet each one significantly steers almost every evaluation format. The strategies recover the same underlying persona while surfacing different facets of it: Dialogue vectors lean toward idealism, Description toward threatened egotism, and all emphasize instrumental evil.

Δ trait expression: vector extracted via row, evaluated on column

Cross-discourse steering works despite dissimilar vectors. Cells: mean Δ (steered − baseline) for the Evil trait; ★ marks FDR-corrected p < 0.05; hover a cell for baseline → steered scores. Chips below: share of extraction generations passing the judge filter per strategy: the strategies differ wildly in yield, yet all produce effective directions.

6It Replicates, and What It Means

On Apertus-8B-2509, a 15T-token multilingual run with a different data mix, tokenizer, and recipe, persona vectors already exist at the earliest public checkpoint (210B tokens, 1.4% of training), transfer to Apertus-Instruct, refine along the same trajectories with less overall drift, and sadism again grows during training. The pattern is not an OLMo quirk.

Persona representations originate in pretraining, not alignment. They are encoded within the first fraction of a percent of training, with no explicit supervision toward them, and survive every alignment stage with post-training only tuning their volume. The assistant persona that emerges in chat-tuned models is a configuration of pretraining-formed traits, so if you want to shape what personas a model can express, pretraining, especially its earliest stretch, is the leverage point.

Citation

arXiv:2605.13329

epfl-dlab/pretraining_persona

@misc{moskvoretskii2026tracingpersonavectorsllm,
      title={Tracing Persona Vectors Through LLM Pretraining},
      author={Viktor Moskvoretskii and Dominik Glandorf and Jorge Medina Moreira and Tanja K{\"a}ser and Robert West},
      year={2026},
      eprint={2605.13329},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2605.13329},
}

1Extract a Direction, Steer Any Checkpoint

Elicit

Filter & subtract

Steer