Compressed foundation models, particularly compressed large language models (LLMs), are increasingly deployed in real-world applications due to their efficiency advantages. Techniques such as quantization, pruning, and distillation significantly reduce memory consumption and latency, making it possible to run these models on resource-constrained devices, such as edge computing platforms and mobile applications. However, compression can also introduce critical robustness and security vulnerabilities, affecting both model reliability and trustworthiness. Compressed models are often more susceptible to adversarial attacks, hallucinations, jailbreak exploits, and degradation in out-of-distribution generalization, raising concerns about their safe deployment in high-stakes applications such as healthcare, finance, and content moderation.

This workshop places a particular emphasis on the robustness challenges of widely deployed compressed foundation models. While existing workshops may discuss general robustness issues in foundation models or focus on efficiency-driven compression, our workshop uniquely targets the intersection of model compression and robustness vulnerabilities. This perspective is crucial because compressed models are now widely used in practice, yet their resilience under real-world threats remains poorly understood and insufficiently addressed. For instance, quantized LLMs are more vulnerable to adversarial prompts, while pruned models may suffer from increased susceptibility to factual errors and hallucinations. Addressing these challenges is not just a theoretical exercise but a pressing need for ensuring secure and reliable deployment.