Provable Adversarial Robustness in In-Context Learning

Abstract

Large language models adapt to new tasks through in-context learning (ICL) without parameter updates. Current theoretical explanations for this capability assume test tasks are drawn from a distribution similar to that seen during pretraining. This assumption overlooks adversarial distribution shifts that threaten real-world reliability. To address this gap, we introduce a distributionally robust meta-learning framework that provides worst-case performance guarantees for ICL under Wasserstein-based distribution shifts. Focusing on linear self-attention Transformers, we derive a non-asymptotic bound linking adversarial perturbation strength ($\rho$), model capacity ($m$), and the number of in-context examples ($N$). The analysis reveals that model robustness scales with the square root of its capacity ($\rho_{\text{max}} \propto \sqrt{m}$), while adversarial settings impose a sample complexity penalty proportional to the square of the perturbation magnitude ($N_\rho - N_0 \propto \rho^2$). Experiments on synthetic tasks confirm these scaling laws. These findings advance the theoretical understanding of ICL's limits under adversarial conditions and suggest that model capacity serves as a fundamental resource for distributional robustness.