Saved in:
Bibliographic Details
Main Authors: Li, Xiuyu, Zhang, Jinkai, Yi, Mingyang, Li, Yu, Wang, Longqiang, Wang, Yue, Fan, Ju
Format: Preprint
Published: 2026
Subjects:
Online Access:https://arxiv.org/abs/2601.21484
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866910235603501056
author Li, Xiuyu
Zhang, Jinkai
Yi, Mingyang
Li, Yu
Wang, Longqiang
Wang, Yue
Fan, Ju
author_facet Li, Xiuyu
Zhang, Jinkai
Yi, Mingyang
Li, Yu
Wang, Longqiang
Wang, Yue
Fan, Ju
contents Reinforcement Learning (RL) post-training alignment for language models is effective, but also costly and unstable in practice, owing to its complicated training process. To address this, we propose a training-free inference method to sample directly from the optimal RL policy. The transition probability applied to Masked Language Modeling (MLM) consists of a reference policy model and an energy term. Based on this, our algorithm, Energy-Guided Test-Time Scaling (ETS), estimates the key energy term via online Monte Carlo, with a provable convergence rate. Moreover, to ensure practical efficiency, ETS leverages modern acceleration frameworks alongside tailored importance sampling estimators, substantially reducing inference latency while provably preserving sampling quality. Experiments on MLM (including autoregressive models and diffusion language models) across reasoning, coding, and science benchmarks show that our ETS consistently improves generation quality, validating its effectiveness and design. The code is available at https://github.com/sheriyuo/ETS.
format Preprint
id arxiv_https___arxiv_org_abs_2601_21484
institution arXiv
publishDate 2026
record_format arxiv
spellingShingle ETS: Energy-Guided Test-Time Scaling for Training-Free RL Alignment
Li, Xiuyu
Zhang, Jinkai
Yi, Mingyang
Li, Yu
Wang, Longqiang
Wang, Yue
Fan, Ju
Machine Learning
Reinforcement Learning (RL) post-training alignment for language models is effective, but also costly and unstable in practice, owing to its complicated training process. To address this, we propose a training-free inference method to sample directly from the optimal RL policy. The transition probability applied to Masked Language Modeling (MLM) consists of a reference policy model and an energy term. Based on this, our algorithm, Energy-Guided Test-Time Scaling (ETS), estimates the key energy term via online Monte Carlo, with a provable convergence rate. Moreover, to ensure practical efficiency, ETS leverages modern acceleration frameworks alongside tailored importance sampling estimators, substantially reducing inference latency while provably preserving sampling quality. Experiments on MLM (including autoregressive models and diffusion language models) across reasoning, coding, and science benchmarks show that our ETS consistently improves generation quality, validating its effectiveness and design. The code is available at https://github.com/sheriyuo/ETS.
title ETS: Energy-Guided Test-Time Scaling for Training-Free RL Alignment
topic Machine Learning
url https://arxiv.org/abs/2601.21484