📄 中文摘要
大型语言模型（LLM）驱动的智能体能够将用户的高级意图转化为环境中的计划和行动。然而，当用户观察到某个结果后，可能会产生疑问：如果我当初的意图表达方式不同，结果会怎样？一个框架被提出，旨在使LLM驱动的自主控制场景能够进行此类反事实推理，同时提供形式化的可靠性保证。该方法对用户、基于LLM的智能体与环境之间的闭环交互进行建模，并设计了反事实生成机制。具体来说，系统分析了原始意图、智能体执行的动作序列以及最终观测到的结果。通过识别导致特定结果的关键决策点和潜在的替代路径，框架能够生成一系列替代意图表述，这些表述在相同或相似的初始条件下，可能会引导智能体采取不同的行动，并最终产生不同的结果。
📄 English Summary
Should I Have Expressed a Different Intent? Counterfactual Generation for LLM-Based Autonomous Control
Large language model (LLM)-powered agents translate high-level user intents into plans and actions within an environment. Following an observed outcome, users often ponder: What if my intent had been articulated differently? A framework is introduced to enable such counterfactual reasoning in agentic LLM-driven control scenarios, while concurrently providing formal reliability guarantees. This approach models the closed-loop interaction among a user, an LLM-based agent, and the environment, designing mechanisms for counterfactual generation. Specifically, the system analyzes the original intent, the agent's executed action sequence, and the final observed outcome. By identifying critical decision points and potential alternative paths that led to a specific result, the framework generates a set of alternative intent phrasings. These alternative phrasings, under identical or similar initial conditions, could have guided the agent to undertake different actions and ultimately produce distinct outcomes. The generation process considers semantic similarity of intents, feasibility within the action space, and reachability within the objective space. To ensure reliability, the framework integrates formal verification techniques, quantifying the predictability of agent behavior and the determinism of outcomes under various intent articulations. This involves modeling the LLM's behavioral patterns given different intent inputs and evaluating their impact on environmental states. Through this methodology, users gain insight into the causality of existing outcomes and can explore potential “what-if” scenarios, thereby optimizing future intent expression strategies and enhancing the efficiency of LLM agents in complex environments, alongside improving user satisfaction.