Search papers, labs, and topics across Lattice.
This paper addresses the challenge of belief inertia in embodied agents, where agents cling to outdated internal beliefs despite new environmental feedback. The authors introduce the Estimate-Verify-Update (EVU) mechanism, which allows agents to predict outcomes, verify them against observations, and update their beliefs accordingly. Experimental results across three benchmarks show that EVU significantly improves task success rates and effectively mitigates belief inertia, enhancing the decision-making capabilities of embodied agents.
Active belief intervention can drastically improve embodied agents' decision-making by overcoming the pitfalls of belief inertia.
Recent advancements in large language models (LLMs) have enabled agents to tackle complex embodied tasks through environmental interaction. However, these agents still make suboptimal decisions and perform ineffective actions, as they often overlook critical environmental feedback that differs from their internal beliefs. Through a formal probing analysis, we characterize this as belief inertia, a phenomenon where agents stubbornly adhere to prior beliefs despite explicit observations. To address this, we advocate active belief intervention, moving from passive understanding to active management. We introduce the Estimate-Verify-Update (EVU) mechanism, which empowers agents to predict expected outcomes, verify them against observations through explicit reasoning, and actively update prior beliefs based on the verification evidence. EVU is designed as a unified intervention mechanism that generates textual belief states explicitly, and can be integrated into both prompting-based and training-based agent reasoning methods. Extensive experiments across three embodied benchmarks demonstrate that EVU consistently yields substantial gains in task success rates. Further analyses validate that our approach effectively mitigates belief inertia, advancing the development of more robust embodied agents. Our code is available at https://github.com/WangHanLinHenry/EVU.
