大语言模型智能体中推理在长程规划中失效的规划中心分析

Large language model (LLM)-based agents demonstrate strong step-by-step reasoning over short horizons, yet frequently fail to maintain coherent behavior across extended planning horizons. This failure stems from a fundamental mismatch: the inherent step-wise reasoning paradigm induces a form of step-wise greedy policy that is adequate for short-term tasks but proves insufficient for long-horizon planning, where initial actions must proactively account for delayed consequences. From a planning-centric perspective, the core challenge for LLM agents in long-range decision-making lies in the discrepancy between their intrinsic sequential reasoning mechanism and the global optimization required for complex planning problems. LLM agents tend to select locally optimal solutions at each step, often neglecting the impact of these choices on future states and the ultimate goal. This myopic behavior is particularly evident in tasks demanding multi-step coordination and trade-offs. For instance, in tasks requiring a specific sequence of operations to achieve a final objective, planning can fail if early actions do not establish the necessary groundwork for subsequent steps, even if each local choice appears reasonable. The agents' lack of foresight into future states and their limited capacity to model long-chain causal relationships hinder their ability to effectively evaluate the long-term value of different action sequences. Furthermore, LLM agents are often susceptible to 'model hallucinations,' generating action sequences that appear plausible but are actually inconsistent with environmental constraints or logical coherence, further exacerbating planning failures.