Visual Dependence Does Not Imply Answerability: A Systematic Analysis of Answerability Signals in Frozen InstructBLIP
cs.CV, cs.AI
Submitted: 2026-10-13
Updated: 2026-10-13
Comments: 10 pages, 3 figures, 5 tables
License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
Sources
- Knowing When Not to Answer: Evaluating Abstention in Multimodal Reasoning Systems
- Qwen2.5-VL Technical Report
- An Overview of Multi-Task Learning in Deep Neural Networks
- InternVL3: Exploring Advanced Training and Test-Time Recipes for Open-Source 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