Mind the Spike: Mechanisms and Brittleness of Visual Massive Activations in Large Vision-Language Models
Visual activation spikes in vision-language models are brittle and can be created or removed with tiny image perturbations.
The study examines massive visual activations across 25 adapter-based large vision-language models built on 18 text-only bases from 10 families, ranging from 2B to 72B parameters. Some models do not form visual spikes; others spike at different rates, typically in deeper layers, on tokens that share least with the rest of the image. Common corruptions often create or move spikes. A trigger-guided attack creates or removes spikes with an l-infinity budget of 1/255 in nine of ten spiking models, and removing only the trigger component largely eliminates spikes while leaving other image tokens nearly unchanged.
- The study covers 25 LVLMs on 18 bases from 10 families, 2B to 72B.
- Eventual spike tokens tend to share least with the rest of the image.
- A 1/255 perturbation creates or removes spikes in nine of ten models.
- Stripping the trigger component before spikes form largely eliminates them.
From the original article
1 min read at arxiv.orgLarge vision-language models (LVLMs) inherit massive activations from their text-only bases: spikes where a few fixed hidden channels receive values thousands of times above the typical magnitude. The text spike systematically appears in early layers at a fixed initial position, independently of input content.
By Jonas Ngnawé · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.