Same Answer, Different Representations: Hidden instability in VLMs
Preprint
Abstract
The robustness of Vision Language Models (VLMs) is commonly assessed through output-level invariance, implicitly assuming that stable predictions reflect stable multimodal processing. In this work, we argue that this assumption is insufficient. We introduce a representation-aware and frequency-aware evaluation framework that measures internal embedding drift, spectral sensitivity, and structural smoothness (spatial consistency of vision tokens), alongside standard label-based metrics. Applying this framework to modern VLMs across the SEEDBench, MMMU, and POPE datasets reveals three distinct failure modes. First, models frequently preserve predicted answers while undergoing substantial internal representation drift; for perturbations such as text overlays, this drift approaches the magnitude of inter-image variability, indicating that representations move to regions typically occupied by unrelated inputs despite unchanged outputs. Second, robustness does not improve with scale; larger models achieve higher accuracy but exhibit equal or greater sensitivity, consistent with sharper yet more fragile decision boundaries. Third, we find that perturbations affect tasks differently: they harm reasoning when they disrupt how models combine coarse and fine visual cues, but on the hallucination benchmarks, they can reduce false positives by making models generate more conservative answers.Keywords
All Related Versions
This publication has 11 references indexed in Scilit:
- Wavelet-Based Mechanistic Interpretability of Vision Transformers via Frequency-Aware Ablations2025
- Visual Robustness Benchmark for Visual Question Answering (VQA)2025
- Calibrating LLM Confidence by Probing Perturbed Representation Stability2025
- Instability in Downstream Task Performance During LLM Pretraining2025
- Response Wide Shut? Surprising Observations in Basic Vision Language Model Capabilities2025
- VLMEvalKit: An Open-Source ToolKit for Evaluating Large Multi-Modality Models2024
- Enhancing Continual Learning in Visual Question Answering with Modality-Aware Feature Distillation2024
- Evaluating Object Hallucination in Large Vision-Language Models2023
- Vision Transformers Are Robust LearnersProceedings of the AAAI Conference on Artificial Intelligence, 2022
- The importance of phase in signalsProceedings of the IEEE, 1981