What Do Visual-Head Discovery Scores Actually Find? Attention Source, Necessity, and Random-Control Pitfalls in Vision-Language Models
cs.CV
Submitted: 2026-10-09
Updated: 2026-10-09
License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
Terminology
Sources
- Gaze Heads: How VLMs Look at What They Describe
- Localizing Model Behavior with Path Patching
- What Do VLMs NOTICE? A Mechanistic Interpretability Pipeline for Gaussian-Noise-free Text-Image Corruption and Evaluation
- How to use and interpret activation patching
- Interpreting Attention Heads for Image-to-Text Information Flow in Large Vision-Language Models
- AtP*: An efficient and scalable method for localizing LLM behaviour to components
- Microsoft COCO: Common Objects in Context
- Transformer Circuit Faithfulness Metrics are not Robust
- MIB: A Mechanistic Interpretability Benchmark
- In-context Learning and Induction Heads
- Targeting the Attention Heads Behind Object Hallucination in LLaVA
- Mechanistic Insights into Functional Sparsity in Multimodal LLMs via CoRe Heads
- Interpretability in the Wild: a Circuit for Indirect Object Identification in GPT-2 small
- Spatial457: A Diagnostic Benchmark for 6D Spatial Reasoning of Large Multimodal Models
Related papers
- Loss Knows Best: Detecting Annotation Errors in Videos via Loss Trajectories
- AnchorWeave: World-Consistent Video Generation with Retrieved Local Spatial Memories
- Benchmarking the Robustness of Foundation Models for Mammography under Domain Shift
- MambaX-Net: Dual-Input Mamba-Enhanced Cross-Attention Network for Longitudinal MRI Segmentation
- TeleOCR: Navigating Document Parsing Across Digital and Camera-Captured Documents
- A Survey on Efficient Vision-Language-Action Models