HaReCAP: Reducing LLM Calls in Long-Horizon Embodied Tasks
A recent preprint on arXiv (2608.16447) presents HaReCAP, a technique aimed at minimizing unnecessary calls to large language models (LLMs) in long-horizon embodied tasks. This method builds upon the ReCAP framework, which employs recursive context management for decomposing tasks at multiple levels and refining parent nodes. HaReCAP tackles the issue of 'last-mile grounding redundancy' by identifying common leaf decisions from successful paths and creating offline, auditable, and abstainable one-step leaf-reflex rules. During execution, it avoids invoking the leaf LLM when a rule is confidently applicable, reducing token and call overhead. The authors, whose names are not mentioned in the abstract, highlight its relevance for AI agents in physical settings like robotics, emphasizing its low-intrusion nature. While experimental results or benchmarks are not detailed, the mechanism hints at possible enhancements in computational efficiency for recursive LLM agents.
Key facts
- HaReCAP is a low-intrusion leaf grounding extension for ReCAP.
- It extracts frequent leaf decisions from successful trajectories.
- Rules are compiled offline into auditable and abstainable one-step leaf-reflex rules.
- At runtime, it skips the leaf LLM call only when a rule confidently applies.
- The method targets last-mile grounding redundancy in long-horizon embodied tasks.
- The paper is available on arXiv with ID 2608.16447.
- The approach aims to reduce LLM-call and token overhead.
- The work builds on the ReCAP framework for recursive context management.
Entities
Institutions
- arXiv