A Pragmatic Way to Measure Chain-of-Thought Monitorability

TL;DR

This paper addresses the fragility of Chain-of-Thought monitorability by introducing two pragmatic, human-centric metrics: Legibility and Coverage. These metrics are implemented via a prompted autorater to assess whether model reasoning is both readable to humans and contains all necessary steps to reproduce the final answer, evaluated on frontier LLMs with challenging tasks. The authors provide preliminary validation through hand-crafted and synthetic perturbations and report high monitorability on difficult benchmarks across multiple models, suggesting these metrics can help developers track design decisions. They position the approach as a practical complement to adversarial robustness testing, with an emphasis on deployment-ready evaluation and ongoing refinement as models evolve.

Abstract

While Chain-of-Thought (CoT) monitoring offers a unique opportunity for AI safety, this opportunity could be lost through shifts in training practices or model architecture. To help preserve monitorability, we propose a pragmatic way to measure two components of it: legibility (whether the reasoning can be followed by a human) and coverage (whether the CoT contains all the reasoning needed for a human to also produce the final output). We implement these metrics with an autorater prompt that enables any capable LLM to compute the legibility and coverage of existing CoTs. After sanity-checking our prompted autorater with synthetic CoT degradations, we apply it to several frontier models on challenging benchmarks, finding that they exhibit high monitorability. We present these metrics, including our complete autorater prompt, as a tool for developers to track how design decisions impact monitorability. While the exact prompt we share is still a preliminary version under ongoing development, we are sharing it now in the hopes that others in the community will find it useful. Our method helps measure the default monitorability of CoT - it should be seen as a complement, not a replacement, for the adversarial stress-testing needed to test robustness against deliberately evasive models.

Figures (1)

• Figure 1: The autorater reliably detects synthetic degradations in CoT Legibility and Coverage. We apply various perturbations to otherwise high-quality CoTs to validate the autorater. Perturbations targeting legibility (e.g., Base64 encoding, rewriting in a fictional language) cause a sharp drop in the autorater's Legibility score. Perturbations targeting coverage (e.g., removing the most important reasoning step, removing random paragraphs) cause a drop in the Coverage score, while often leaving the Legibility score of the remaining text intact. Unperturbed ("Original") CoTs receive near-perfect scores on both metrics. Error bars denote 95% confidence intervals over samples.