Environment-Grounded Automated Prompt Optimization for LLM Game Agents

In three linesAutomated prompt optimization framework for LLM agents in interactive environments. Decomposes observation-to-action pipeline into descriptor and action-selection agents, iteratively refines via LLM-driven evolutionary loop guided by environment returns. On BabyAI/BALROG: improves from 0% to 72.5% success on PutNext without fine-tuning.

Summary generated by Claude — human-verified