BadWAM: Right Mind but Wrong Action
Qi Li · hf · 2026-07-17
BadWAM: World-Action Models Can "Think Right, Act Wrong"
This paper investigates the adversarial vulnerability of World-Action Models (WAMs). WAMs couple action generation with future world prediction, making them often perceived as more robust, interpretable, and safe; however, the authors point out that this "imagination-execution consistency" is actually quite fragile.
Proposed Attacks
BadWAM uniformly describes and evaluates a class of World-Action Drift Attacks against WAMs:
- Action-only attack: Prioritizes disrupting the task by directly pushing the model towards failed actions
- Imagination-preserving attack: Pursues stealth, making the model still predict plausible futures on the surface while actions drift during execution
Results
- Significantly reduces closed-loop execution success rates across different WAM variants
- In the provided examples, the action-only attack dropped the success rate from 96.5% to 43.1%
- Even with some future-grounding regularization, strong attack effects can still be maintained
Conclusion: A WAM "thinking correctly" does not equate to "acting safely".
More from Embodied
- Robot learning researcher joins Mila to work on physical AI and RL — pcastr · 2026-10-27
- Full teardown: inside Zmotor's 32mm linear actuator for humanoids — MarwaEldiwiny · 2026-10-27
- Full Teardown of a Zmotor Linear Actuator Used in China's Humanoid Robots — MarwaEldiwiny · 2026-10-27
- World Model Assigns Weight and Friction to Objects, Then Simulates with a Physics Engine — jonstephens85 · 2026-10-27
- First unboxing shots of Apple's foldable "iPhone Duo" surface — chrisfirst · 2026-10-27
- Bedrock Robotics Raises $600M Series C as Autonomous Excavators Do Paid Work on US Job Sites — davidpattersonx · 2026-10-27