A new paper proposes letting AI learn to write like a specific scientist from just a handful of their old abstracts - no fine-tuning marathon required.
Researchers tested three ways to transfer an author's writing style to a language model using only a few example abstracts per person. The first, contrastive activation steering, nudges the model's internal activations toward an author's style while holding the topic fixed, comparing real abstracts against style-neutral versions of the same content. The second trains a small network to predict those steering vectors automatically, skipping the need for a predefined list of style traits. The third, and best performing, is a hypernetwork that predicts custom LoRA adapter weights for each author on the fly.
Full fine-tuning captured style well but wrecked the model's fluency - a predictable trade-off, but a useful warning for anyone building personalized writing tools. The hypernetwork approach held onto both style accuracy and output quality, and it worked on authors it had never seen before, which matters more than performance on authors used in training.
The researchers also found that hand-picked and machine-predicted steering vectors point in almost unrelated directions yet work equally well - a reminder that there's rarely one 'correct' way to capture someone's voice, mechanically or otherwise.