New Method Boosts LLM Agent Efficiency for Complex Tasks

Large language model (LLM) agents often struggle with long, complex interactive tasks because they typically operate by making one decision and executing one primitive action per LLM round. While this allows for frequent replanning, it becomes highly inefficient when many routine action sequences are required. Attempts to train agents to emit variable-length action chunks using standard reinforcement learning have largely failed, leading to either overly short or excessively long, inflexible action sequences. The core problem identified is the agents' inability to learn effective chunk boundaries. To solve this, researchers propose SPACE (Skill-Guided Adaptive Action Chunking). SPACE addresses this by distilling chunk-boundary supervision directly from programmatic skills derived from successful task trajectories. This temporal structure is then integrated into a primitive-chunk policy through a hybrid on-/off-policy optimization approach with chunk-aware credit assignment. Experiments on challenging environments like ALFWorld and ScienceWorld demonstrate that SPACE significantly improves success rates (by 7.0%-31.3%) while drastically reducing the average number of LLM decision rounds (up to 78.9%), making LLM agents much more efficient for long-horizon tasks.