Persona drift becomes a measurement
Finetuning changes a model's character has long been an observation people could make about a model's outputs after the fact. This theme is about turning that observation into a number computed from the model's own activations, checkable before a training run even finishes and, in one version, before it starts at all.
The page links five pieces into that measurement: a core metric that predicts how strongly a trait will show up once training ends, a generalized version of the same metric that shows related traits moving together, an explanation of why that co-movement accounts for narrow training producing broad misbehavior, an earlier forecast computed purely from the training data itself, and a final framing that treats deliberately induced and accidental persona shifts as the same measurable process.
Persona-vectors (2025) turns an anecdote -- finetuning changes a model's character -- into a quantitative object built from five linked pieces. Finetuning shift supplies the core metric, predicting post-finetuning values of trait expression score at r=0.76-0.97, a genuine validation because trait expression score comes independently from GPT-4.1-mini reading generated text, not from any activation-space quantity. Cross-trait persona correlation analysis generalizes the same construction across trait pairs, projecting one trait's finetuning shift against another trait's behavioral change and finding a real but weaker signal (r=0.34-0.86) -- evidence trait directions carry trait-specific signal rather than a generic drift alarm, even though the negative traits and humor move together and against optimism. That correlation structure is what finally gives Betley et al.'s emergent misalignment a mechanistic account: because trait directions are not orthogonal, training pressure that raises one trait's expression predictably raises correlated traits too, explaining why narrow training produces broad misbehavior. Projection difference forecasts finetuning shift's magnitude before any training happens, computed purely from pre-finetuning training data, though the forecast is uneven: strong for evil and sycophancy on both base models, weak for hallucination on Qwen (r=0.408, barely significant) and still the weakest trait on Llama (r=0.593). Finally, finetuning shift supplies a trait-agnostic predictor for emergent misalignment itself, computed identically whether a dataset was built to elicit a trait or not, collapsing "designed" and "emergent" persona shifts into one measurable underlying process.