AMARIS: A Memory-Augmented Rubric Improvement System for Rubric-Based Reinforcement Learning

In three linesAMARIS enhances rubric-based RL by integrating persistent evaluation memory. The system accumulates evaluation diagnostics over time, retrieves them via static and semantic search, and continuously adapts reward rubrics. Experiments show performance gains with ~5% time overhead.

Summary generated by Claude — human-verified