Frontier Models Can Take Actions at Low Probabilities

TL;DR

Frontier models demonstrate a surprising ability to perform actions at very low probabilities with high calibration, especially when external entropy or explicit reasoning is involved, raising concerns for model oversight.

Contribution

This study evaluates the capability of frontier models to take low-probability actions and highlights their potential to evade detection during pre-deployment evaluations.

Findings

Models maintain high calibration at rates below 1 in 100,000 with entropy.

Larger models perform better at low-rate calibration when given target rates.

Explicit Chain-of-Thought reasoning is crucial for successful low-rate actions.

Abstract

Pre-deployment evaluations inspect only a limited sample of model actions. A malicious model seeking to evade oversight could exploit this by randomizing when to "defect": misbehaving so rarely that no malicious actions are observed during evaluation, but often enough that they occur eventually in deployment. But this requires taking actions at very low rates, while maintaining calibration. Are frontier models even capable of that? We prompt the GPT-5, Claude-4.5 and Qwen-3 families to take a target action at low probabilities (e.g. 0.01%), either given directly or requiring derivation, and evaluate their calibration (i.e. whether they perform the target action roughly 1 in 10,000 times when resampling). We find that frontier models are surprisingly good at this task. If there is a source of entropy in-context (such as a UUID), they maintain high calibration at rates lower than 1 in 100,000 actions. Without external entropy, some models can still reach rates lower than 1 in 10,000. When target rates are given, larger models achieve good calibration at lower rates. Yet, when models must derive the optimal target rate themselves, all models fail to achieve calibration without entropy or hint to generate it. Successful low-rate strategies require explicit Chain-of-Thought (CoT) reasoning, so malicious models attempting this approach could currently be caught by a CoT monitor. However, scaling trends suggest future evaluations may be unable to rely on models' lack of target rate calibration, especially if CoT is no longer legible.