A new training recipe folds a teacher model's token-by-token hints into the reward math that grades AI systems on verified answers.
Researchers describe UECR-GRPO, short for Unified Entropy-Calibrated Credit Redistribution for GRPO, a method for training math-reasoning models that combines two techniques already in use. Reinforcement learning with verifiable rewards checks only whether a final answer is correct, so it gives models almost no signal about which tokens along the way helped or hurt. On-policy distillation supplies that denser, token-level feedback by comparing a student's output to a teacher model, but a teacher's preferences do not always line up with correctness. Earlier attempts to merge the two let teacher signals in only after the verifier's scores were already normalized, or reweighted tokens in ways that could silently change the total credit a response received.
UECR-GRPO blends the teacher score into the same objective as the verifier reward before that normalization step, then redistributes credit token by token based on the gap between the teacher's and the model's own prior predictions, dialing back the adjustment whenever the teacher itself looks uncertain. Across five math benchmarks with Qwen3 models, it beat the strongest existing baseline by 0.89 and 0.56 percentage points at the 1.7B and 4B parameter sizes.
Those are not the kind of numbers that justify a press release, and this paper is not trying to sell one. It is a reminder that most current progress in training AI to reason through math problems is incremental credit-assignment bookkeeping, not new capability.