Qi Li, Xingyi Yang, Xinchao Wang
A newer breed of robot brain does two things at once: it decides what action to take and it imagines how the world will look afterward. People like this because it seems to come with a built-in safety check. If the robot's mental picture of the future looks fine, surely the action it chose is fine too. This paper shows that trust is misplaced.
BadWAM is a recipe for adversarial attacks that quietly pry apart what the robot imagines from what it actually does, using only small tweaks to the camera image. One version goes for maximum damage, pushing the robot toward task-failing actions, dropping success from 96.5% to 43.1% in one case. The nastier version keeps the imagined future looking normal while the executed action drifts off, so the very safety check people counted on stays green while the robot misbehaves.
This is a summary of the abstract, so see the paper for the attack details and defenses.
World-action models (WAMs) are emerging as a promising foundation for embodied control: rather than predicting actions alone, they learn representations that couple action generation with future world prediction. This coupling is often viewed as a source of robustness, interpretability, and safety, as a robot's action can in principle be checked against its imagined future. In this paper, we show that this assumption is fragile. We introduce BadWAM, a unified framework for…
T^2MLR: Transformer with Temporal Middle-Layer Recurrence
arXiv (cs.AI) · July 16, 2026Benchmarking Multimodal Large Language Models for Scientific Visualization Literacy
arXiv (cs.CL) · July 16, 2026Linear representations of grammaticality in neural language models
arXiv (cs.AI) · July 16, 2026MedFailBench: A Clinician-Built Open-Source Benchmark for Medical AI Safety Boundary Inspection