Chong Li

Chloe Li, Mary Phuong, Noah Y. Siegel

Trustworthy evaluations of dangerous capabilities are increasingly crucial for determining whether an AI system is safe to deploy. One empirically demonstrated threat is sandbagging - the strategic underperformance on evaluations by AI models or their developers. A promising defense is to monitor a model's chain-of-thought (CoT) reasoning, as this could reveal its intentions and plans. In this work, we measure the ability of models to sandbag on dangerous capability evaluations against a CoT monitor by prompting them to sandbag while being either monitor-oblivious or monitor-aware. We show that both frontier models and small open-sourced models can covertly sandbag against CoT monitoring 0-shot without hints. However, they cannot yet do so reliably: they bypass the monitor 16-36% of the time when monitor-aware, conditioned on sandbagging successfully. We qualitatively analyzed the uncaught CoTs to understand why the monitor failed. We reveal a rich attack surface for CoT monitoring and contribute five covert sandbagging policies generated by models. These results inform potential failure modes of CoT monitoring and may help build more diverse sandbagging model organisms.

Xin Li, Yaohua Hu, Chong Li et al.

In this paper, we discuss the statistical properties of the $\ell_q$ optimization methods $(0<q\leq 1)$, including the $\ell_q$ minimization method and the $\ell_q$ regularization method, for estimating a sparse parameter from noisy observations in high-dimensional linear regression with either a deterministic or random design. For this purpose, we introduce a general $q$-restricted eigenvalue condition (REC) and provide its sufficient conditions in terms of several widely-used regularity conditions such as sparse eigenvalue condition, restricted isometry property, and mutual incoherence property. By virtue of the $q$-REC, we exhibit the stable recovery property of the $\ell_q$ optimization methods for either deterministic or random designs by showing that the $\ell_2$ recovery bound $O(ε^2)$ for the $\ell_q$ minimization method and the oracle inequality and $\ell_2$ recovery bound $O(λ^{\frac{2}{2-q}}s)$ for the $\ell_q$ regularization method hold respectively with high probability. The results in this paper are nonasymptotic and only assume the weak $q$-REC. The preliminary numerical results verify the established statistical property and demonstrate the advantages of the $\ell_q$ regularization method over some existing sparse optimization methods.