Turns out an AI grader that reasons before scoring can match one trained on 100× more data — just by thinking.

This explores whether process reward models — the systems that grade each step of an AI's reasoning — get better *and* cheaper to train when you let them think before they judge. The corpus gives an unusually clean answer: yes, and the data savings are dramatic. Reframing step-scoring as a generative task, where the judge writes out chain-of-thought reasoning before assigning a score, beats the older discriminative approach (a model that just outputs a number) by orders of magnitude in label efficiency. A 1.5B-parameter generative PRM beats GPT-4o, and ThinkPRM matches full-dataset verifiers using just 1% of the labels Can generative reasoning beat discriminative models with less training data?. The striking part isn't only that reasoning helps accuracy — it's that reasoning substitutes for data.

What's happening is a convergence that several teams found independently. Adding a reasoning trace before scoring lets the reward model spend more compute at evaluation time, the same test-time scaling trick that makes reasoning models strong — except pointed at judgment instead of answering Can reward models benefit from reasoning before scoring?. StepWiser pushes this further: instead of classifying whether a step is good, the judge *meta-reasons about the reasoning* — asking why a step works — which lifts both accuracy and data efficiency Can judges that reason about reasoning outperform classifier rewards?. The pattern is consistent across GenPRM, ThinkPRM, and RM-R1: reasoning-before-judging raises the ceiling that outcome-only scoring can't reach.

But the interesting frontier is where reasoning-before-judging *isn't enough on its own.* A judge that reasons about a polished answer is one thing; a judge that reasons about a real thinking trace — full of backtracking, dead ends, and branching — is another. Standard PRMs actually degrade on those messy traces, and ReasonFlux-PRM has to explicitly supervise both the trajectory and the final response, treating failed exploration as informative rather than as an error Why do standard process reward models fail on thinking traces?. So the reasoning that makes a PRM data-efficient also has to be trained to tolerate the untidy shape of real reasoning.

Two adjacent findings sharpen what "efficient" is really buying you. Domain matters: a general reasoning PRM misses factual and regulatory errors that a knowledge-grounded, finance-specific PRM catches, because reasoning about *logical* coherence isn't the same as reasoning about *correctness* Can general process reward models catch factual errors in finance?. And the source of the signal matters: process rewards mined from what search agents read-but-don't-cite structurally block reward hacking, capturing intermediate quality without the model learning to fabricate it Can search agent behavior yield reliable process rewards for reasoning?.

Here's the thing you didn't know you wanted to know: the label savings may come from the same well as the reasoning gains themselves. Work on RLVR shows that reinforcement mostly *activates* strategies already latent in a pretrained model rather than teaching new ones — a single example can suffice to switch on a capability What does reward learning actually do to model reasoning?. A generative PRM that reasons before judging is plausibly doing the same thing: eliciting an evaluation skill the base model already has, which is exactly why it needs 1% of the labels instead of the whole dataset. Reasoning isn't just a better output format — it's a cheaper way to reach a capability that was already there.