Weightless Fine-Tuning Personalizes LLMs Without Weight Updates

Supervised fine-tuning (SFT) is a common method for adapting Large Language Models (LLMs) to specific tasks or user preferences. However, SFT becomes computationally expensive and resource-intensive in scenarios requiring individual personalization, as each user would need a separate model instance, leading to high optimization, storage, and retraining costs. A new technique called Weightless Fine-Tuning (WFT) has been proposed to overcome these limitations. WFT is a training-free method that operates during decoding, approximating the distributional impact of SFT without altering the model's weights. It achieves this by computing supervised residuals on an author's training sequence and then transporting these residuals to the current prompt using a cross-prefix transport operator. This operator, estimated from dropout-induced cross-covariance, effectively replaces gradient-based parameter updates with logit-space corrections. Experiments on three personalization benchmarks demonstrated that WFT achieved superior average performance, matching or exceeding SFT on individual tasks and outperforming other lightweight baselines. Crucially, WFT approached SFT performance using less than 7% of the effective computation, indicating a highly efficient personalization approach. Logit-level analysis confirmed a high cosine similarity between the logit shifts induced by WFT and SFT, suggesting WFT effectively captures the desired distributional adaptation.