Evil persona vector saturates by approximately 12B tokens in OLMo-3

Quantitative result showing Evil emerges earliest due to ubiquity and simplicity in pretraining data

Source paper

extracted_from

Tracing Persona Vectors Through LLM Pretraining

(2026) · Viktor Moskvoretskii · Dominik Glandorf · Jorge Medina Moreira · Tanja Käser +1

Neighborhood — ranked by edge-count

Papers (1)

paper

Tracing Persona Vectors Through LLM Pretraining
mentions

Related by similarity (8)

cosine ≥ 0.65 · no typed edge

Entities in the same semantic neighborhood but without a typed relation to this one — candidates for new edges or unrecognized duplicates.

Persona vectors form within 0.22% of OLMo-3 pretraining (≈12.6B tokens)finding0.840
Core quantitative result answering RQ1 for OLMo-3
Evil persona steering delta of +9.69 (p=0.011) at stage1-step3000 (12.6B tokens) in OLMo-3finding0.833
First significant evil persona steering result; establishes emergence at 0.22% pretraining
Finetuning shift along evil persona vector correlates r=0.826 (Qwen) and r=0.930 (Llama) with evil trait expression scorefinding0.793
Quantitative result for evil trait showing persona vector prediction power on both model architectures
Sadism facet of Evil persona grows at the beginning of Evil persona development in OLMo-3 pretrainingfinding0.765
Facet-level finding correlating geometric refinement with qualitative semantic change
Decomposing the evil persona vector via SAE reveals features including insulting language (F12061, s=0.336, TES=91.2), deliberate cruelty (F128289, s=0.306, TES=84.9), and malicious code (F14739, s=0.334, TES=78.6)finding0.764
SAE decomposition reveals interpretable fine-grained features composing the evil persona vector
Chen et al. 2025 identified seven persona vectors in Qwen2.5-7B-Instruct and Llama-3.1-8B-Instruct: evil, sycophancy, hallucination, optimism, impoliteness, apathy, and humorfinding0.760
Systematic identification of multiple coexisting persona vectors in two open-source models
Evil and Impolite persona vectors evolve in parallel in MDS space, suggesting intertwined representationsfinding0.759
Geometric finding from MDS analysis suggesting shared representational structure
On Llama-3.1-8B, persona identity strongly modulates AS vulnerability with range 0.087-0.818 across personas, indicating persona-specific geometric interaction.finding0.758
Contrast with Gemma/Qwen showing Llama-specific persona-AS interaction