Rethinking RL Evaluation: Can Benchmarks Truly Reveal Failures of RL Methods?
This work addresses the critical problem of unreliable evaluation in RL for AI researchers, highlighting that current benchmarks are insufficient for assessing generalization, which is incremental in proposing diagnostic tools and principles rather than a new method.
The paper tackles the problem that current reinforcement learning (RL) benchmarks for large language models are inadequate for evaluating progress, as training on training sets achieves nearly the same performance as on test sets, failing to reveal generalization failures. It introduces a diagnostic suite and Oracle Performance Gap metric to quantify this issue, finding that RL methods struggle with distribution shifts, difficulty variations, and counterfactual scenarios despite strong benchmark scores.
Current benchmarks are inadequate for evaluating progress in reinforcement learning (RL) for large language models (LLMs).Despite recent benchmark gains reported for RL, we find that training on these benchmarks' training sets achieves nearly the same performance as training directly on the test sets, suggesting that the benchmarks cannot reliably separate further progress.To study this phenomenon, we introduce a diagnostic suite and the Oracle Performance Gap (OPG) metric that quantifies the performance difference between training on the train split versus the test split of a benchmark. We further analyze this phenomenon with stress tests and find that, despite strong benchmark scores, existing RL methods struggle to generalize across distribution shifts, varying levels of difficulty, and counterfactual scenarios: shortcomings that current benchmarks fail to reveal.We conclude that current benchmarks are insufficient for evaluating generalization and propose three core principles for designing more faithful benchmarks: sufficient difficulty, balanced evaluation, and distributional robustness.