ETS: Energy-Guided Test-Time Scaling for Training-Free RL Alignment

News

[2026.05.01] ETS is accepted as a poster at ICML 2026.
[2026.01.29] We have uploaded ETS paper to arXiv, and open-sourced ETS on Github!

New features in this repo

[2026.05.15] Support vllm for ETS and AIME24 avg@32 evaluation for all methods.

Introduction

We introduce ETS (Energy-Guided Test-Time Scaling), a training-free inference method that samples directly from the optimal RL policy under a unified Masked Language Modeling (MLM) framework that covers both:

Autoregressive Models (ARMs)
Diffusion Language Models (DLMs)

Core idea:
For RL objective, the optimal policy admits a closed-form structure. ETS leverages this to construct an optimal transition kernel that factorizes into:

a reference transition given by a base model $p_{\mathrm{ref}}$, and
an energy term that is a conditional expectation of exponentiated rewards.

Setup

Run the following script to setup environment.

git clone https://github.com/sheriyuo/ETS.git
cd ETS
pip install -e .

Evaluation

Hyperparameters you will tune

ETS compute is dominated by three hyperparameters:

$M$: number of candidates per guidance step
$K$: number of Monte Carlo estimation, $K=3$ works best in most cases
$I$: number of guidance steps

For evaluating autoregressive models (Qwen), the ETS compute parameters map to:

$M$: m_candidates
$K$: k_monte_carlo
$I$ is implicit . It is determined by the total decoding length and block granularity:
- max_length = total generation length $d_x$
- block_size = block length $B$
- so $I = \lceil \mathrm{max_length} / \mathrm{block_size} \rceil$

For evaluating diffusion language models (LLaDA), the mapping is explicit:

$I$: guide_steps
$M$: num_candidates
$K$: monte_carlo_num

We evaluate in a pass@1 setting on:

Math/Reasoning: MATH500, GSM8K
Coding: HumanEval
STEM: GPQA (Diamond)

Autoregressive model

Evaluate ETS with transformers:

cd qwen
bash eval.sh

Evaluate ETS with vllm:

cd qwen_vllm
bash eval.sh

Diffusion Language model

cd llada
bash eval.sh

Evaluation for AIME24 (lm_eval)

Download BytedTsinghua-SIA/AIME-2024 to a local directory and update the data path in aime24/aime24.yaml to point to your local dataset.
Replace the existing utils.py and aime24.yaml in lm_eval/tasks/aime/. For example:

rm -rf /usr/local/miniconda3/lib/python3.10/site-packages/lm_eval/tasks/aime/aime24.yaml
rm -rf /usr/local/miniconda3/lib/python3.10/site-packages/lm_eval/tasks/aime/utils.py
cp aime24/aime24.yaml /usr/local/miniconda3/lib/python3.10/site-packages/lm_eval/tasks/aime/aime24.yaml
cp aime24/utils.py /usr/local/miniconda3/lib/python3.10/site-packages/lm_eval/tasks/aime/utils.py

Execute eval_aime.sh.

Citation

@inproceedings{li2026ets,
  title={ETS: Energy-Guided Test-Time Scaling for Training-Free RL Alignment},
  author={Li, Xiuyu and Zhang, Jinkai and Yi, Mingyang and Li, Yu and Wang, Longqiang and Wang, Yue and Fan, Ju},
  booktitle={Forty-third International Conference on Machine Learning},
  year={2026}
}

Shuaijie She