Self-promo
A COLM 2026 poster explores 'weightless fine-tuning' for personalizing LLMs
Its presenter says per-user supervised fine-tuning gets expensive when tailoring a model to thousands or millions of people.
TLDR
In an October 8 invitation to a COLM 2026 poster session, a presenter described “weightless fine-tuning” as an approach to personalizing LLMs without actual fine-tuning. They said supervised fine-tuning is a strong option for one user but gets expensive at the scale of thousands or millions: it requires separate optimization and model or LoRA weights for each user, plus retraining as new data arrives.
Combined views
300
2 Sources, first seen ago
