Amazon-Microsoft paper: skill-guided action chunking lifts agent success to 67.2% while halving LLM calls
rohanpaul_ai · x · 2026-09-04
A new Amazon-Microsoft paper argues long-horizon LLM agents shouldn't call the model after every tiny action — the hard part is learning which actions can safely run together.
Their method SPACE:
- Induces two-level programmatic skills from successful trajectories, using subskill boundaries as supervision for meaningful action chunks
- Distills this temporal structure into a policy that emits variable-length primitive actions, with no skill library needed at test time
On ScienceWorld, success rose from 35.9% to 67.2% and average LLM rounds fell from 10.2 to 5.2; the abstract reports 7.0%-31.3% gains over the strongest baselines on ALFWorld and ScienceWorld with up to 78.9% fewer decision rounds. Takeaway: don't reconsider every step, and don't blindly batch actions — train agents to know when to keep acting and when to look again.
Related event: SPACE Cuts Agent LLM Calls by 78.9% While Boosting Success(3 posts)→
More from coding & agent
- PostgreSQL 19 slips to November as LLM-assisted code review floods committers with edge cases — sanjaykalra · 2026-10-21
- LangChain adds Human-in-the-Loop beta to Managed Deep Agents — LangChain · 2026-10-21
- Voxtral 3 ships a LiveKit-based voice agent SDK streaming listen-think-speak — MistralAI · 2026-10-21
- Open-sourced pixel art normals generator ships as Rust CLI and wasm library — HugoDuprez · 2026-10-20
- Eval researcher teases guide to subjective-task evals, with a curated resource list — cwolferesearch · 2026-10-20
- Eon cuts engineer onboarding from 2 days to 2 minutes with an MCP Gateway on Bedrock AgentCore — AWS ML Blog · 2026-10-20