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)→
More from coding & agent
- Junie Local runs a coding agent on your Mac with unlimited usage, no credits — heyronir · 2026-09-04
- POSTECH's PACE uses coordinated agents to surface hidden conflicts in user requests — POSTECH · 2026-09-04
- Astra solves Excel World Championship cases ~4x faster than human champions using pure computer use — sandersted · 2026-09-04
- Dev builds a Magicka-inspired MMO in just 3 days using fable 5.1 — TAbrodi · 2026-09-04
- Creator assigns overnight tasks to AI agents, hopes for renders not a coup — bilawalsidhu · 2026-09-04
- Amazon-Microsoft paper: skill-guided action chunking lifts agent success to 67.2% while halving LLM calls — rohanpaul_ai · 2026-09-04