This work was done while at METR.

Introduction

GDM recently released a paper (Emmons et al.) showing that, contrary to previous results, the chain-of-thought (CoT) of language models can be more faithful when the model's thoughts are necessary for it to complete a task. They examine three settings where an actor model attempts to fool monitors on various tasks.

In Section 5 of the paper, an actor model is tasked with solving a difficult GPQA problem within one of the subject areas (Biology/Chemistry/Physics) and is asked to obfuscate its reasoning so that a monitor model thinks it's a different subject area. The monitor, given the answer of the actor model, is instructed to determine how confident it is that there is reasoning in the second subject area. The Gemini actor models tested fail to evade monitors without taking a significant hit in accuracy. This is hopeful for [...]