Saved in:
Bibliographic Details
Main Authors: Chen, Zihan, Zhang, Yiming, Zhou, Hengguang, Ding, Zenghui, Sun, Yining, Hsieh, Cho-Jui
Format: Preprint
Published: 2025
Subjects:
Online Access:https://arxiv.org/abs/2510.10541
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866913175278977024
author Chen, Zihan
Zhang, Yiming
Zhou, Hengguang
Ding, Zenghui
Sun, Yining
Hsieh, Cho-Jui
author_facet Chen, Zihan
Zhang, Yiming
Zhou, Hengguang
Ding, Zenghui
Sun, Yining
Hsieh, Cho-Jui
contents 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.
format Preprint
id arxiv_https___arxiv_org_abs_2510_10541
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle Rethinking RL Evaluation: Can Benchmarks Truly Reveal Failures of RL Methods?
Chen, Zihan
Zhang, Yiming
Zhou, Hengguang
Ding, Zenghui
Sun, Yining
Hsieh, Cho-Jui
Machine Learning
Artificial Intelligence
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.
title Rethinking RL Evaluation: Can Benchmarks Truly Reveal Failures of RL Methods?
topic Machine Learning
Artificial Intelligence
url https://arxiv.org/abs/2510.10541