SPACE cuts agent LLM calls by 78.9% while raising success rate on long-horizon tasks
dair_ai · x · 2026-09-04
SPACE tackles a core inefficiency of ReAct-style agents: one primitive action per model round means most of a long-horizon episode is spent re-deciding routine sequences.
- Training variable-length action chunks with vanilla RL fails because the policy never learns where a chunk should end.
- SPACE induces two-level programmatic skills from successful trajectories and uses subskill boundaries as direct chunk-boundary labels.
- It distills this temporal structure via hybrid on/off-policy optimization with chunk-aware credit assignment.
Result: 78.9% fewer LLM calls with higher success rate.
More from coding & agent
- Perplexity says Astra excels at computer use, coming to Comet browser — AravSrinivas · 2026-09-04
- Utopia: open-source enterprise world model built on a bitemporal knowledge graph — techNmak · 2026-09-04
- OpenAI showcases GPT-6 Astra: 3D London history, matcha site and DEF CON puzzle via parallel agents — OpenAI · 2026-09-04
- swyx Burned 20B Tokens Stress-Testing Astra on Real AI Engineering Tasks — All for Under $6/Hour — charliermarsh · 2026-09-04
- LangChain launches free LangSmith Essentials course covering the full agent dev lifecycle in 60 minutes — Hacubu · 2026-09-04
- OpenAI's Lukasz Kaiser: even we can't pinpoint what caused the Christmas coding-agent jump — a_karvonen · 2026-09-04