MemReward Improves RL Fine-Tuning With Graph Rewards
Researchers introduce MemReward, a graph-based experience memory framework that stores LLM rollouts as heterogeneous graph nodes and uses a GNN to propagate sparse reward labels during reinforcement learning fine-tuning; authors submitted the paper to arXiv on March 13, 2026. Experiments on Qwen2.5-3B and 1.5B across mathematics, question answering, and code generation show MemReward reaches 97.3% of Oracle performance with 20% labels on 3B and 96.6% on 1.5B, scaling to 99.4% at 70% labels and surpassing Oracle on out-of-domain tasks.
Key Points
- 1Stores rollouts as heterogeneous graph nodes linking queries, thinking processes, and answers with similarity edges
- 2Trains a GNN on labeled nodes to propagate rewards to unlabeled rollouts for label efficiency
- 3Achieves ~97% Oracle performance with 20% labels and scales to 99.4% at 70% labels
Scoring Rationale
High novelty and broad applicability across LLM tasks, limited by being an arXiv preprint without peer review.
Sources
Public references used for this report.
Practice interview problems based on real data
1,625 SQL & Python problems across 15 industry datasets — the exact type of data you work with.
Try 250 free problems
