Cue-GRPO: Rarity-Aware Credit Redistribution Improves RLVR Performance

A recent paper on arXiv (2608.03467v2) presents Cue-GRPO, a technique designed to tackle credit concentration issues in reinforcement learning with verifiable rewards (RLVR). The authors define a challenge termed 'multiplicity-induced structure-level credit concentration,' where in GRPO, frequent correct solution patterns gather positive coefficient mass based on their sampling rates, while infrequent patterns receive minimal credit. To address this, they introduce a partition-conditioned rule that reallocates positive advantages based on the rarity of clusters. Cue-GRPO applies this rule without the need for auxiliary-model inference, utilizing deterministic Strategy Cues to create rollout-local partitions of verified-correct traces. Experiments conducted on Qwen2.5-Math-7B and Llama-3.1-8B-Instruct demonstrate that Cue-GRPO enhances AIME repeated-sampling performance, particularly with higher sampling budgets. The paper also explores Credit Redistribution (CR) under Judge Partitions (JP) as an additional extension.