Induced Numerical Instability: Hidden Costs in Multimodal Large Language Models

TL;DR

This paper reveals a new failure mode in multimodal large language models caused by numerical instability, demonstrating that small input changes can significantly degrade performance, which is not detected by traditional adversarial methods.

Contribution

The study introduces a novel method to induce numerical instability in multimodal models, exposing a hidden failure mode that impacts model robustness and performance.

Findings

Performance degrades significantly with minimal input changes.

Numerical instability can be intentionally induced to cause model failure.

This failure mode is distinct from traditional adversarial attacks.

Abstract

The use of multimodal large language models has become widespread, and as such the study of these models and their failure points has become of utmost importance. We study a novel mode of failure that causes degradation in performance indirectly by optimizing a loss term that seeks to maximize numerical instability in the inference stage of these models. We apply this loss term as the optimization target to construct images that, when used on multimodal large language models, cause significant degradation in the output. We validate our hypothesis on state of the art models large vision language models (LLaVa-v1.5-7B, Idefics3-8B, SmolVLM-2B-Instruct) against standard datasets (Flickr30k, MMVet, TextVQA, VQAv2, POPE, COCO) and show that performance degrades significantly, even with a very small change to the input image, compared to baselines. Our results uncover a fundamentally different vector of performance degradation, highlighting a failure mode not captured by adversarial perturbations.