New method cuts LLM personalization costs by 93%
Researchers propose 'weightless fine-tuning' to personalize LLMs without weight updates
University of Michigan researchers have unveiled Weightless Fine-Tuning (WFT), a novel approach to personalize large language models (LLMs) without the computational overhead of traditional fine-tuning. Published in arXiv (2608.11342), the method computes supervised residuals on a user's training data and transports them to new prompts using a cross-prefix transport operator, eliminating the need for gradient-based weight updates.
In benchmark tests on three LaMP personalization tasks, WFT achieved the best average performance across datasets and matched or exceeded standard supervised fine-tuning (SFT) results on individual tasks. Crucially, it accomplished this with less than 7% of the effective computation required by SFT, while maintaining a 0.875 cosine similarity between logit shifts induced by WFT and SFT across 95% of next-token probability mass. This suggests WFT can deliver similar personalization benefits without model weight modifications or retraining cycles.
- WFT achieves 87.5% cosine similarity to SFT's logit shifts while using <7% of SFT's computation
- Tested on LaMP benchmarks, WFT outperforms other lightweight personalization methods
- Method eliminates need for weight updates by operating in logit-space with dropout-induced transport operators
Why It Matters
Enables cost-effective LLM personalization for enterprises without expensive retraining infrastructure