When VLMs Destroy What They Know: Spatial Representations Fall Below Lexical Priors
cs.CL
Submitted: 2026-10-09
Updated: 2026-10-09
Code: https://github.com/Shubhf/CASP-VL
License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
Sources
- MMCOMPOSITION: Revisiting the Compositionality of Pre-trained Vision-Language Models
- CS-CLIP: Compositional Scene Graph-guided CLIP for Robust Compositional Reasoning
- Qwen2.5 Technical Report
- The information bottleneck method
- Same Answer, Different Representations: Hidden instability in VLMs
- Exploring Information Processing in Large Language Models: Insights from Information Bottleneck Theory
Related papers
- Exploring Solution Divergence and Its Effect on Large Language Model Problem Solving
- Ishigaki-IDS-Bench: A Benchmark for Generating Information Delivery Specification from BIM Information Requirements
- Subliminal Steering: Stronger Encoding of Hidden Signals
- MedStruct-S: A Benchmark for Key Discovery, Key-Conditioned QA and Semi-Structured Extraction from OCR Clinical Reports
- The End of Transformers? On Challenging Attention and the Rise of Sub-Quadratic Architectures
- Untangling the Mechanisms of Misleading Context in Medical Question Answering