GR-Ben: New Benchmark for Evaluating Process Reward Models Beyond Math

Researchers introduced GR-Ben, a process-level benchmark for evaluating process reward models (PRMs) across two primary reasoning domains—science and logic—and nine subdomains. Existing benchmarks focus mainly on mathematical reasoning, failing to assess PRMs' error detection in diverse scenarios. GR-Ben tests 22 models including PRMs and large language models (LLMs), revealing that error-detection ability declines in non-mathematical domains. The benchmark addresses the need for comprehensive evaluation as LLMs often produce flawed intermediate reasoning steps. The paper is available on arXiv.