MINT：用于目标驱动知识差距推理和主动启发式学习的最小信息神经符号树

Joint planning via language-based interactions is a pivotal aspect of human-AI teaming, especially in open-world scenarios. These real-world planning challenges frequently encounter incomplete information and various unknowns, such as the specific objects involved or the precise human goals and intentions. Such informational voids inevitably lead to significant knowledge gaps within the joint planning process. Addressing this, a critical problem involves discovering optimal interaction strategies for AI agents to actively elicit necessary human inputs in object-driven planning contexts. To tackle this, the Minimal Information Neuro-Symbolic Tree (MINT) framework is introduced. MINT is designed to reason about the profound impact of these knowledge gaps. Furthermore, it leverages a self-play mechanism with MINT to systematically optimize the AI agent's elicitation strategies and the formulation of its queries. Specifically, MINT constructs a symbolic tree by generating propositions of potential human-AI interactions. This structured approach allows for a systematic exploration of interaction possibilities, enabling the AI agent to strategically identify and acquire missing information from human collaborators. The core objective is to enhance the AI agent's ability to navigate uncertainty and incompleteness, leading to more robust and effective joint planning outcomes.