Researchers have found a way to reach inside a large language model and turn up — or down — its agreeableness, conscientiousness, or other personality dimensions without retraining the whole thing.
The paper, posted to arXiv, proposes using sparse autoencoders and contrastive activation analysis to locate directions in a model's residual stream that correspond to each of the five OCEAN personality traits. From there, the researchers add a small shift to the model's hidden states — a "steering vector" — that nudges the target trait in a chosen direction. To find the right size and mix of shifts, they ran a grid search over linear weightings, balancing how much personality expression they got against how much benchmark performance they gave up.
Most prior work on shaping LLM personality relied on prompt engineering (telling the model to "act extroverted") or fine-tuning (retraining it to be so). Both are blunt instruments: prompts can be overridden mid-conversation, and fine-tuning is expensive and hard to reverse. A steering vector applied at the activation level is lighter and more surgical — closer to editing a setting than rewriting the software.
The catch, which the authors acknowledge, is that grid search optimization over feature combinations does not scale gracefully as models grow larger and personality dimensions multiply. It is also worth noting that the OCEAN framework itself is a psychological model with known limitations — mapping its five traits cleanly onto token prediction is an assumption the field has not fully stress-tested.