@PyTorch
What happens when an AI agent recognizes that an action is unsafe, but executes it anyway? In this clip from PyTorch Conference Europe 2026, Besmira Nushi, Senior Manager, AI Research at NVIDIA, examines the “action-thought disconnect,” a failure mode in which a model’s expressed reasoning does not match the action it takes. Using an example from computer use agent evaluation, Nushi explains why task-completion scores alone may miss critical failures and argues for evaluating the full execution trajectory. 🔗 Watch the full keynote: https://t.co/Tow3g4PGlm #PyTorchCon