Sandbag Detection through Model Impairment

Cameron Tice, Philipp Alexander Kreer, Nathan Helm-Burger, Prithviraj Singh Shahani, Fedor Ryzhenkov, Teun van der Weij, Felix Hofstätter, Jacob Haimes

Track: Technical

Keywords: AI Safety, AI Governance, Sandbagging, Capability Evaluations, Noise Injection.

TL;DR: This paper demonstrates that noise injection can be used to effectively detect sandbagging behaviors in LLMs.

Abstract: Capability evaluations play a critical role in ensuring the safe deployment of frontier AI systems, but this role may be undermined by intentional underperformance or "sandbagging." We present a novel model-agnostic method for detecting sandbagging behavior using noise injection. Our approach is founded on the observation that introducing Gaussian noise into the weights of models either prompted or fine-tuned to sandbag can considerably improve their performance. We test this technique across a range of model sizes and multiple-choice question benchmarks (MMLU, AI2, WMDP). Our results demonstrate that noise injected sandbagging models show performance improvements compared to standard models. Leveraging this effect, we develop a classifier that consistently identifies sandbagging behavior. Our unsupervised technique can be immediately implemented by frontier labs or regulatory bodies with access to weights to improve the trustworthiness of capability evaluations.

Submission Number: 114