Brilliant paper on long-horizon agents.

They cut 78.9% of an agent's LLM calls while raising its success rate.

Here is how:

It turns out that ReAct issues one primitive action per model round. That allows frequent replanning, and on long-horizon tasks it spends most of the episode re-deciding routine sequences that were never in doubt.

Training an agent to emit variable-length action chunks with standard RL fails because the policy never learns where a chunk should end, so it either falls back to single actions or commits to sequences that run far too long.

SPACE derives the supervision from data it already has.

It induces two-level programmatic skills from successful trajectories and uses the subskill boundaries as direct chunk-boundary labels, then distills the temporal structure into a primitive-chunk policy with hybrid on-policy and off-policy optimization and chunk-aware credit assignment.

On ALFWorld and ScienceWorld it improves success rates by 7.0 to 31.3% over the strongest baseline in each setting while reducing average LLM decision rounds by up to 78.9%.

Paper: https://arxiv.org/abs/2609.02042

Chat with Paper: https://academy.dair.ai/papers/act-more-decide-less-skill-guided-adaptive-action-chunking-for-long-horizon-llm-2609.02042