Multimodal AI models hide internal instability despite correct answers

Original: When Correct Decisions Hide Internal Stress: Decision-State Probing in Multimodal Language Models

Source: arxiv.org ↗