Paper page - BadWAM: When World-Action Models Dream Right but Act Wrong | Buobe
Paper page - BadWAM: When World-Action Models Dream Right but Act Wrong
Buobe IA context · why it matters
A new class of WAM-specific attacks is introduced — This highlights a previously unaddressed vulnerability in world-action models.
Two criteria for evaluating these attacks are defined: attack strength and stealthiness — Researchers can now systematically assess how effective and covert such attacks might be.
Fique de olho
Attackers may exploit this to manipulate WAMs, potentially leading to unsafe robot behaviors.
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.
