SAGA: Scene-Aware, Goal-Evolving Agents for Long-Horizon Strategy Game Planning

arXiv cs.AI · 6 hours ago

Researchers developed SAGA, an LLM multi-agent framework designed for long-horizon strategic planning in complex games like FreeCiv. The system achieved the highest mean final score on the CivRealm benchmark compared to six other methods and reduced output tokens by 27%. SAGA addresses token overflow and planning coordination issues through semantic scene graphs, domain-specific controllers, and intermediate goal-setting mechanisms that improve performance across successive games.