SPACE (EMNLP 2026): learning chunk boundaries from trajectories cuts agent LLM rounds by up to 78.9%

rohanpaul_ai · x · 2026-09-04

The arXiv paper behind rohanpaul's thread (EMNLP 2026 camera-ready) details SPACE. ReAct-style one-action-per-round wastes LLM rounds on long-horizon tasks, while naive RL for variable-length action chunks collapses to single actions or over-commits — both from failing to learn chunk boundaries.

SPACE induces two-level programmatic skills from successful trajectories, uses subskill boundaries as chunk-boundary supervision, and distills the temporal structure into a primitive-chunk policy via hybrid on-/off-policy optimization with chunk-aware credit assignment. On ALFWorld and ScienceWorld it beats the strongest baselines by 7.0%-31.3% success and cuts decision rounds by up to 78.9%.

Related event: SPACE Cuts Agent LLM Calls by 78.9% While Boosting Success(3 posts)→

Original post →

More from coding & agent

coding & agent channel →