VLMs exhibit only slight performance degradation on hallucination benchmarks when substantial image tokens are removed, with layer-wise analysis showing increased visual token similarity in deeper layers, suggesting current benchmarks inadequately test fine-grained visual grounding.
VLind-bench: Measuring language priors in large vision- language models
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.CV 1years
2026 1verdicts
UNVERDICTED 1representative citing papers
citing papers explorer
-
Seeing without Looking: Do Vision-Language Benchmarks Really Test Vision?
VLMs exhibit only slight performance degradation on hallucination benchmarks when substantial image tokens are removed, with layer-wise analysis showing increased visual token similarity in deeper layers, suggesting current benchmarks inadequately test fine-grained visual grounding.