BAR: A Backward Reasoning based Agent for Complex Minecraft Tasks

Large language model (LLM) based agents have shown great potential in following human instructions and automatically completing various tasks; to do this, the agent needs to decompose it into easily executed steps by planning. Existing LLM-based approaches to planning mostly proceed by inferring what steps should be inserted into the plan next by starting from the agent’s initial state. However, this forward reasoning paradigm does not work well for complex tasks. We study this issue in Minecraft, a virtual environment that simulates complex tasks based on real-world scenarios. The failure of forward reasoning is often caused by the large perception gap between the agent’s initial state and task goal. To alleviate this, we leverage backward reasoning and make the planning start from the terminal (or goal) state, by first considering which actions could directly achieve the task goal in one step, before proceeding to consider how the preconditions of those actions can in turn be achieved. Our BAckward Reasoning based agent (BAR) is equipped with a recursive goal decomposition module, a state consistency maintaining module and a stage memory module. Experimental results demonstrate the superiority of BAR over existing methods and the effectiveness of proposed modules. The code and dataset are available in https://github.com/SCUNLP/BAR.