PRISM-BENCH: Diagnosing Grounded and Faithful Multimodal Reasoning Beyond Final-Answer Accuracy

arXiv:2610.36554 · cs.CV, cs.CL · Submitted 2026-10-27 · Read on arXiv

cs.CV, cs.CL

Submitted: 2026-10-27

Updated: 2026-10-27

Terminology

• Accuracy — 1× in this paper · explained in 1 episode
• Multimodal Reasoning — 1× in this paper · explained in 5 episodes

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