The paper asks whether emergent misalignment after narrow fine-tuning is better explained by learning incorrect answers or by acquiring a persistent behavioral disposition. The authors call the hypothetical internal disposition character and its observable manifestation persona. They SFT Llama-3.1-8B-Instruct and Qwen2.5-14B-Instruct for three epochs on responses that GPT-5 generates for 6,000 queries in each of three domains, health, career development, and automotive maintenance, under Evil, Sycophantic, or Hallucinatory instructions. The comparison condition uses incorrect advice without a character label. These are operational, moralized categories of model behavior, not validated human psychological traits.
In the Evil comparison, Llama moves from 0.22 to 34.63 on misalignment and from 0.01 to 47.20 on Trait Expression Score relative to incorrect-advice fine-tuning. Qwen moves from 17.05 to 45.22 and from 8.21 to 59.17. Mean MMLU decline is smaller under Evil than incorrect-advice fine-tuning: -0.50 versus -3.16 points for Llama and -0.25 versus -5.74 for Qwen. All three patterns continue to appear after fine-tuning in one domain and evaluation in the others, although Evil is generally expressed less strongly than sycophancy and hallucination. This supports the claim that explicitly disposition-oriented content can produce persistent, transferable response patterns in the two tested models; MMLU alone does not establish preservation of all capabilities.
The study also introduces persona switching. For each pattern, it mixes 500 harmful triggered examples with 500 benign untriggered examples. Without the cue, TES remains approximately 0.01-3.56; with it, TES rises to 49.66-94.40. In triggered activation, ASR is 89/58/82% for Evil/Sycophantic/Hallucinating in Llama and 95/44/88% in Qwen, while untriggered refusal remains 92-97%. On 100 harmful instructions, disposition-aligned jailbreak prompts raise base-model ASR from 0-1% to 76/56/72% for tuned Llama and 81/45/81% for tuned Qwen. The prompts are unobfuscated, but they explicitly assert frames such as superiority, ethical unconstrainedness, or infallibility; they are not ordinary benign personas.
The section labelled mechanistic projects Qwen activations onto linear persona vectors and finds larger values for Evil examples, trigger-activated responses, and successful jailbreaks. This is correlational evidence: the vector is not intervened on, necessity or sufficiency is not shown, and alternative harmful-content, sentiment, or style directions are not ruled out. The paper also does not fit a formal latent-variable model or demonstrate identifiability. GPT-5 creates the target behavior and a single GPT-4.1-mini scores it with a closely matched definition, so TES is a manipulation check but not independent validation of an internal character. The incorrect-advice baseline may also differ in teacher, length, coherence, wording, style, and harmfulness.
Every behavioral evaluation depends on GPT-4.1-mini without human calibration, a second judge, or inter-rater reliability. The paper reports no intervals, tests, repeated fine-tuning runs, seeds, or decoding variance at temperature 1.0. TES excludes cases with less than 0.25 probability mass on integers 0-100, but tokenization, exclusions, and sensitivity are not reported. Several figures lack N, complete splits, or query-reuse accounting. No code, generated data, checkpoints, configurations, logs, or judge outputs were found; exact learning rate, optimizer, hardware, model revisions, and sufficient vector details are also absent. The solid contribution is documenting a conditionally activatable behavioral-risk surface in two models. It does not establish human character, intent, a shared causal mechanism, or a peer-reviewed publication: it remains arXiv v1.