Minimal neuron ablation triggers catastrophic collapse in the language core of Large Vision-Language Models

2025-11-30

AI Summary

This study reveals a critical structural vulnerability in large vision-language models (LVLMs): the catastrophic failure of language understanding upon the ablation of merely a few key neurons. To address this, we propose Critical Neuron Analysis (CAN), a method that integrates neuron-level progressive masking with persistent activation tracking to precisely identify cross-layer critical neurons. Experiments on mainstream LVLMs—including LLaVA and InstructBLIP—demonstrate that the projection layer within the language module is especially vulnerable, and performance collapse follows a distinct two-stage pattern. Strikingly, masking just four critical neurons induces a precipitous drop in multimodal understanding performance. This work constitutes the first systematic identification of “single-point-failure”-style fragile units in LVLMs, establishing a novel, interpretable paradigm for robustness evaluation and safety enhancement of multimodal foundation models.

Technology Category

• Computer Vision: Large Vision Models
• Machine Learning: Large Multimodal Models (LMMs)
• Natural Language Processing: Safety and Robustness

Application Category

• User Modeling, Personalization and Recommendation: Large Language Models (LLM) for user modeling and recommendation
• Graph Algorithms and Modeling for the Web: Foundation models and LLMs for Web-related graphs
• Search and Retrieval-Augmented AI: Large language models for search

Abstract

Large Vision-Language Models (LVLMs) have shown impressive multimodal understanding capabilities, yet their robustness is poorly understood. In this paper, we investigate the structural vulnerabilities of LVLMs to identify any critical neurons whose removal triggers catastrophic collapse. In this context, we propose CAN, a method to detect Consistently Activated Neurons and to locate critical neurons by progressive masking. Experiments on LLaVA-1.5-7b-hf and InstructBLIP-Vicuna-7b reveal that masking only a tiny portion of the language model's feed-forward networks (just as few as four neurons in extreme cases) suffices to trigger catastrophic collapse. Notably, critical neurons are predominantly localized in the language model rather than in the vision components, and the down-projection layer is a particularly vulnerable structure. We also observe a consistent two-stage collapse pattern: initial expressive degradation followed by sudden, complete collapse. Our findings provide important insights for safety research in LVLMs.

Problem

• Identify critical neurons causing catastrophic collapse in LVLMs
• Propose a method to detect consistently activated neurons
• Analyze structural vulnerabilities in language components of LVLMs

Innovation

• Detect critical neurons via progressive masking
• Localize vulnerabilities in language model components
• Identify two-stage collapse pattern in LVLMs