Process Reward Agents for Steering Knowledge-Intensive Reasoning

TL;DR

This paper introduces Process Reward Agents (PRA), a method for online, step-wise rewards in reasoning tasks, improving accuracy and generalization without retraining the underlying policy models.

Contribution

PRA enables dynamic, online reward signals during inference, outperforming prior post hoc methods and generalizing across various frozen policy models in knowledge-intensive reasoning.

Findings

Achieves 80.8% accuracy on MedQA with Qwen3-4B, setting a new state of the art.

Improves accuracy of frozen models by up to 25.7% without retraining.

Generalizes to unseen models ranging from 0.5B to 8B parameters.

Abstract

Reasoning in knowledge-intensive domains remains challenging as intermediate steps are often not locally verifiable: unlike math or code, evaluating step correctness may require synthesizing clues across large external knowledge sources. As a result, subtle errors can propagate through reasoning traces, potentially never to be detected. Prior work has proposed process reward models (PRMs), including retrieval-augmented variants, but these methods operate post hoc, scoring completed trajectories, which prevents their integration into dynamic inference procedures. Here, we introduce Process Reward Agents (PRA), a test-time method for providing domain-grounded, online, step-wise rewards to a frozen policy. In contrast to prior retrieval-augmented PRMs, PRA enables search-based decoding to rank and prune candidate trajectories at every generation step. Experiments on multiple medical reasoning benchmarks demonstrate that PRA consistently outperforms strong baselines, achieving 80.8% accuracy on MedQA with Qwen3-4B, a new state of the art at the 4B scale. Importantly, PRA generalizes to unseen frozen policy models ranging from 0.5B to 8B parameters, improving their accuracy by up to 25.7% without any policy model updates. More broadly, PRA suggests a paradigm in which frozen reasoners are decoupled from domain-specific reward modules, allowing the deployment of new backbones in complex domains without retraining.