Reinforcement learning (RL) fine-tuning improves large language model (LLM) reasoning using reward signals on sampled rollouts. When reward labels are available for only a subset of rollouts, restricting policy updates to that subset leaves the remaining rollouts unused for policy optimization. Semi-supervised learning propagates labels from labeled to unlabeled samples, but existing methods either diffuse labels with fixed similarity weights or treat each sample independently, ignoring the relations among queries, thinking processes, and answers. We propose MemReward, a graph-based experience memory that propagates rewards to unlabeled rollouts during online policy optimization. MemReward stores queries and the rollouts (thinking processes and final answers) of an initial LLM policy as nodes of a heterogeneous graph connected by similarity and structural edges, and warms up a GNN on labeled rollouts to propagate rewards over this graph. During online RL fine-tuning, unlabeled rollouts are attached to the graph by query similarity, and the policy is updated with ground-truth rewards on labeled rollouts and GNN-predicted rewards on the rest. We evaluate MemReward on 13 benchmarks in mathematics, question answering, and code generation with verifiable rewards, and further with LLM-judge, human-preference, and continuous rewards. With ground-truth rewards on only 20% of rollouts, MemReward achieves 96.6% of the performance of fully supervised training (Oracle) on Qwen2.5-1.5B and 97.3% on Qwen2.5-3B, where it also outperforms label propagation and FixMatch-style self-training. It carries over to Llama backbones up to 8B and closely approaches Oracle on out-of-domain tasks.

翻译：暂无翻译