Drift: Decoding-time Personalized Alignments with Implicit User Preferences

TL;DR

Drift is a training-free framework that personalizes large language models at decoding time by modeling user preferences with interpretable attributes, achieving superior results with minimal data.

Contribution

Introducing Drift, a novel method for decoding-time personalization of LLMs using implicit preferences without extensive training or gradient updates.

Findings

Outperforms RLHF baselines with only 50-100 examples

Efficient and interpretable personalization at decoding time

Effective on both synthetic and real datasets

Abstract

Personalized alignments for individual users have been a long-standing goal in large language models (LLMs). We introduce Drift, a novel framework that personalizes LLMs at decoding time with implicit user preferences. Traditional Reinforcement Learning from Human Feedback (RLHF) requires thousands of annotated examples and expensive gradient updates. In contrast, Drift personalizes LLMs in a training-free manner, using only a few dozen examples to steer a frozen model through efficient preference modeling. Our approach models user preferences as a composition of predefined, interpretable attributes and aligns them at decoding time to enable personalized generation. Experiments on both a synthetic persona dataset (Perspective) and a real human-annotated dataset (PRISM) demonstrate that Drift significantly outperforms RLHF baselines while using only 50-100 examples. Our results and analysis show that Drift is both computationally efficient and interpretable.