New method aims to separate query-relevant reasoning from perceptual noise in vision-language models
Researchers introduce QCVS, a question-contrastive visual steering technique for multimodal reasoning that addresses inference-time steering entangling query-relevant dynamics with low-level perceptual noise.
Read the full story at arXiv cs.CV ↗
Timeline · 1 report
- 2026-11-10 05:00 · arXiv cs.CV
QCVS: Question-Contrastive Visual Steering for Multimodal Reasoning