CLR-Bench: benchmark for self-evolving agents on a hidden-rule curriculum (arXiv:2610.40758v1)
The paper introduces CLR-Bench, a Curriculum of Latent Rules that constructs chains adding one hidden rule per level and releases 20 chains across 10 domains; probes ensure base models don’t already infer the groupings. Six recent self-evolution methods are evaluated: most barely improve (best gain ≤5 points), a fixed explicit lesson raises performance by ~14.5–16.5 points, and with a stronger model (GPT-6-Sol) ACE and Dynamic Cheatsheet reach ~78% accuracy and solve ~59–60% of single cells versus ~53% accuracy and ~8% cell-solve rate for the certifying model.