Search papers, labs, and topics across Lattice.
This paper introduces SPACE, a novel approach for training long-horizon LLM agents that utilizes skill-guided adaptive action chunking to improve efficiency in interactive tasks. By distilling chunk-boundary supervision from programmatic skills derived from successful trajectories, SPACE enables agents to learn when to group actions into variable-length chunks, overcoming the limitations of traditional single-action protocols. Experimental results demonstrate that SPACE significantly enhances success rates by 7.0%-31.3% and reduces decision rounds by up to 78.9% compared to existing methods.
Variable-length action chunks can boost LLM agent performance by up to 31% while slashing decision-making time by nearly 79%.
Large language model (LLM) agents for long-horizon interactive tasks typically follow a ReAct-style protocol, issuing one primitive action per LLM round. While this enables frequent replanning, it is inefficient for long-horizon tasks where many rounds are spent on routine action sequences. A natural alternative is to let the agent emit variable-length action chunks. However, naively training such policies with standard reinforcement learning fails: the agent either collapses to single-action behavior or over-commits to excessively long sequences. Both failures share a common root cause: the inability to learn chunk boundaries. We propose SPACE, which addresses this challenge by distilling chunk-boundary supervision from trajectory-induced programmatic skills. We induce two-level programmatic skills from successful trajectories, where subskill boundaries serve as direct chunk-boundary supervision. This temporal structure is then distilled into a primitive-chunk policy via hybrid on-/off-policy optimization with chunk-aware credit assignment. Experiments on ALFWorld and ScienceWorld show that SPACE improves success rates by 7.0%-31.3% over the strongest baseline in each setting while reducing average LLM decision rounds by up to 78.9%.
