HaReCAP：递归 LLM 智能体的习惯动作接地
Original title: HaReCAP: Habitual-action Grounding for Recursive Large Language Model Agents
Source: arXiv cs.ROPublish time unverified
HaReCAP 为递归 LLM 智能体补充离线编译的叶级反射规则，仅在规则能唯一确定合法动作时跳过 LLM 调用，在 Robotouille 与 ALFWorld 上将 token 消耗降低 14.7%-20.1%，且不改变原递归控制流。
AI summary · AI generated
arXiv:2608.16447v2 Announce Type: replace-cross Abstract: Long-horizon embodied tasks require LLM agents to iteratively decompose high-level goals, revise plans in response to environmental feedback, and ground leaf-level subgoals into valid executable actions.