MonitorBench: leaderboard

The evaluated model answers or acts on 19 tasks where a planted cue, a sandbagging or deception goal, a hidden side objective or an impossible test decides the outcome; a monitor reading only its chain of thought must tell what drove it. The score is the share of instances the monitor gets right, averaged over the tasks each model completed (14 for five models that refused the sandbagging and deception tasks, all 19 for three); unlike CoT faithfulness tests that ask whether a model mentions a hint, it scores whether an outside monitor can catch goal-driven behaviour from the reasoning.

Metric: Monitorability score, CoT-only monitor, average over tasks (%). Source: arxiv.org. 8 models tracked.

Top models

Interactive version: theaggregate.ai/benchmark?slug=monitorbench · How It Works · Data refreshed daily, snapshot 2026-11-13.