English

GenARM: Reward Guided Generation with Autoregressive Reward Model for Test-time Alignment

Computation and Language 2025-07-16 v5

Abstract

Large Language Models (LLMs) exhibit impressive capabilities but require careful alignment with human preferences. Traditional training-time methods finetune LLMs using human preference datasets but incur significant training costs and require repeated training to handle diverse user preferences. Test-time alignment methods address this by using reward models (RMs) to guide frozen LLMs without retraining. However, existing test-time approaches rely on trajectory-level RMs which are designed to evaluate complete responses, making them unsuitable for autoregressive text generation that requires computing next-token rewards from partial responses. To address this, we introduce GenARM, a test-time alignment approach that leverages the Autoregressive Reward Model--a novel reward parametrization designed to predict next-token rewards for efficient and effective autoregressive generation. Theoretically, we demonstrate that this parametrization can provably guide frozen LLMs toward any distribution achievable by traditional RMs within the KL-regularized reinforcement learning framework. Experimental results show that GenARM significantly outperforms prior test-time alignment baselines and matches the performance of training-time methods. Additionally, GenARM enables efficient weak-to-strong guidance, aligning larger LLMs with smaller RMs without the high costs of training larger models. Furthermore, GenARM supports multi-objective alignment, allowing real-time trade-offs between preference dimensions and catering to diverse user preferences without retraining. Our project page is available at: https://genarm.github.io.

Keywords

Cite

@article{arxiv.2410.08193,
  title  = {GenARM: Reward Guided Generation with Autoregressive Reward Model for Test-time Alignment},
  author = {Yuancheng Xu and Udari Madhushani Sehwag and Alec Koppel and Sicheng Zhu and Bang An and Furong Huang and Sumitra Ganesh},
  journal= {arXiv preprint arXiv:2410.08193},
  year   = {2025}
}

Comments

Published at the Thirteenth International Conference on Learning Representations (ICLR 2025)

R2 v1 2026-06-28T19:16:45.622Z