GR-Ben: A General Reasoning Benchmark for Evaluating Process Reward Models

2026-05-01

AI Summary

Existing process reward models (PRMs) are primarily confined to mathematical reasoning and lack the capability to detect errors in intermediate steps across diverse, general-purpose reasoning scenarios, further hindered by the absence of cross-domain evaluation benchmarks. To address this gap, this work proposes GR-Ben—the first process-level evaluation benchmark for general reasoning—spanning nine subtasks across scientific and logical domains. Leveraging multi-domain data, the benchmark includes error-annotated samples to systematically evaluate 22 PRMs and large language models (LLMs). Experimental results reveal that current models exhibit substantially limited error-detection performance outside mathematical contexts: PRMs struggle to identify knowledge-related errors, while LLMs underperform in detecting computational mistakes.

Abstract

Currently, process reward models (PRMs) have exhibited remarkable potential for test-time scaling. Since large language models (LLMs) regularly generate flawed intermediate reasoning steps when tackling a broad spectrum of reasoning and decision-making tasks, PRMs are required to possess capabilities for detecting process-level errors in real-world scenarios. However, existing benchmarks primarily focus on mathematical reasoning, thereby failing to comprehensively evaluate the error detection ability of PRMs across diverse reasoning scenarios. To mitigate this gap, we introduce GR-Ben, a process-level benchmark specifically designed for assessing PRM's performance across two primary reasoning domains (science and logic) and nine subdomains. We conduct extensive experiments on a diverse set of 22 models, encompassing both PRMs and LLMs, and derive two key findings: (1) In domains beyond mathematical reasoning, the error-detection ability of existing PRMs and LLMs is found to be markedly weaker by comparison.(2) In general, PRMs are less adept at identifying knowledge-based errors, whereas LLMs exhibit poorer performance in detecting computational errors.We hope GR-Ben can foster future researches on PRMs for general domains, thereby enhancing the reasoning capabilities of LLMs.

Problem

• process reward models
• reasoning benchmark
• error detection
• general reasoning
• large language models

Innovation

• Process Reward Models
• General Reasoning Benchmark
• Error Detection
• Test-Time Scaling
• Large Language Models