Proceedings of the Annual Meeting of the Cognitive Science Society

Enabling embodied agents to complete long-horizon tasks in complex environments remains a fundamental challenge. Current LLM-based agents primarily focus on logical reasoning, yet they lack an intuitive perception of latent environmental risks. This often leads to failure during execution as they inadvertently overlook potential hazards. Inspired by the Somatic Marker Hypothesis in neuroscience, we propose a novel architecture called Cognitive Behavior Tree (CBT). By integrating LLM-based symbolic reasoning within behavior trees with activation steering risk premonition, we empower agents with the dual capacity for both deliberate planning and a premonition of potential risks. Extensive experiments in Minecraft demonstrate that CBT excels at handling high-difficulty tasks, achieving an average success rate of 26.60%—significantly surpassing existing state-of-the-art methods.