MonitorBench: A Comprehensive Benchmark for Chain-of-Thought Monitorability in Large Language Models

TL;DR

MonitorBench is an open-source benchmark designed to evaluate the ability of large language models to reliably monitor their reasoning processes through chains of thought, especially under stress conditions.

Contribution

It introduces a comprehensive, diverse set of test instances and stress-test settings to systematically assess CoT monitorability in LLMs.

Findings

Monitorability is higher when decision-critical factors influence intermediate reasoning.

More capable LLMs tend to have lower monitorability.

Stress-tests can reduce monitorability by up to 30% in some tasks.

Abstract

Large language models (LLMs) can generate chains of thought (CoTs) that are not always causally responsible for their final outputs. When such a mismatch occurs, the CoT no longer faithfully reflects the actual reasons (i.e., decision-critical factors) driving the model's behavior, leading to the reduced CoT monitorability problem. However, a comprehensive and fully open-source benchmark for thoroughly evaluating CoT monitorability remains lacking. To address this gap, we propose MonitorBench, a systematic benchmark for evaluating CoT monitorability in LLMs. MonitorBench provides: (1) a diverse set of 1,514 test instances with carefully designed decision-critical factors across 19 tasks spanning 7 categories to characterize \textit{when} CoTs can be used to monitor the factors driving LLM behavior; and (2) two stress-test settings to quantify \textit{the extent to which} CoT monitorability can be degraded. Extensive experiments across multiple popular LLMs with varying capabilities show that CoT monitorability is higher when the decision-critical factors shape the intermediate reasoning process without merely influencing the final answer. More capable LLMs tend to exhibit lower monitorability. And all evaluated LLMs can intentionally reduce monitorability under stress-tests, with monitorability dropping by up to 30\% in some tasks that do not require structural reasoning over the decision-critical factors. Overall, MonitorBench provides a basis for further research on evaluating future LLMs, studying advanced stress-test monitorability techniques, and developing new monitoring approaches. The code is available at https://github.com/ASTRAL-Group/MonitorBench.