Saved in:
Bibliographic Details
Main Authors: Li, Yafu, Hu, Xuyang, Qu, Xiaoye, Li, Linjie, Cheng, Yu
Format: Preprint
Published: 2025
Subjects:
Online Access:https://arxiv.org/abs/2501.12895
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866929684110901248
author Li, Yafu
Hu, Xuyang
Qu, Xiaoye
Li, Linjie
Cheng, Yu
author_facet Li, Yafu
Hu, Xuyang
Qu, Xiaoye
Li, Linjie
Cheng, Yu
contents Large language models (LLMs) demonstrate impressive performance but lack the flexibility to adapt to human preferences quickly without retraining. In this work, we introduce Test-time Preference Optimization (TPO), a framework that aligns LLM outputs with human preferences during inference, removing the need to update model parameters. Rather than relying on purely numerical rewards, TPO translates reward signals into textual critiques and uses them as textual rewards to iteratively refine its response. Evaluations on benchmarks covering instruction following, preference alignment, safety, and mathematics reveal that TPO progressively improves alignment with human preferences. Notably, after only a few TPO steps, the initially unaligned Llama-3.1-70B-SFT model can surpass the aligned counterpart, Llama-3.1-70B-Instruct. Furthermore, TPO scales efficiently with both the search width and depth during inference. Through case studies, we illustrate how TPO exploits the innate capacity of LLM to interpret and act upon reward signals. Our findings establish TPO as a practical, lightweight alternative for test-time preference optimization, achieving alignment on the fly. Our code is publicly available at https://github.com/yafuly/TPO.
format Preprint
id arxiv_https___arxiv_org_abs_2501_12895
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle Test-Time Preference Optimization: On-the-Fly Alignment via Iterative Textual Feedback
Li, Yafu
Hu, Xuyang
Qu, Xiaoye
Li, Linjie
Cheng, Yu
Computation and Language
Large language models (LLMs) demonstrate impressive performance but lack the flexibility to adapt to human preferences quickly without retraining. In this work, we introduce Test-time Preference Optimization (TPO), a framework that aligns LLM outputs with human preferences during inference, removing the need to update model parameters. Rather than relying on purely numerical rewards, TPO translates reward signals into textual critiques and uses them as textual rewards to iteratively refine its response. Evaluations on benchmarks covering instruction following, preference alignment, safety, and mathematics reveal that TPO progressively improves alignment with human preferences. Notably, after only a few TPO steps, the initially unaligned Llama-3.1-70B-SFT model can surpass the aligned counterpart, Llama-3.1-70B-Instruct. Furthermore, TPO scales efficiently with both the search width and depth during inference. Through case studies, we illustrate how TPO exploits the innate capacity of LLM to interpret and act upon reward signals. Our findings establish TPO as a practical, lightweight alternative for test-time preference optimization, achieving alignment on the fly. Our code is publicly available at https://github.com/yafuly/TPO.
title Test-Time Preference Optimization: On-the-Fly Alignment via Iterative Textual Feedback
topic Computation and Language
url https://arxiv.org/abs/2501.12895