Search papers, labs, and topics across Lattice.
This paper introduces Process Reward Agents (PRA), a novel test-time method that provides online, step-wise rewards to a frozen policy for knowledge-intensive reasoning tasks. PRA leverages domain-grounded rewards to rank and prune candidate trajectories during search-based decoding, enabling dynamic inference. Experiments on medical reasoning benchmarks show that PRA significantly improves accuracy, achieving state-of-the-art performance at the 4B scale and generalizing to unseen frozen policy models.
Unlock up to 25.7% accuracy gains on frozen LLMs in knowledge-intensive domains, without any retraining, by dynamically rewarding reasoning steps.
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.
ETH