Process reward models (PRMs) are widely used to guide test-time scaling in large language models, providing step-level feedback during reasoning. But these models can themselves be unreliable, and the new paper argues that this inaccuracy stems from two distinct issues, though the abstract does not detail them.

To address this, the authors propose a reliability- and conservatism-aware Bayesian belief framework. Rather than treating process rewards as fixed scores, the method models them as beliefs with uncertainty, aiming to avoid overconfident or misleading signals during scaling.

The work is positioned as a step toward more trustworthy process supervision. Because the abstract is brief, the precise mechanisms and experimental results are not yet available, but the framing suggests a shift from point estimates to uncertainty-aware reward modeling.