Preprint reports gains from grounding game-coaching AI in live state

Researchers report that rules tying an AI game coach to current match and session data raised its turn-level factual accuracy to 96.7%, from 61.1% for a prompting baseline and 69.8% for another agent design. Session-level accuracy rose to 83.5%, from 20.0% and 26.5%, respectively, in a 200-session benchmark. Their approach places structured checks around the language model so its advice follows the player's current state. The results come from one in-game coaching setting, and the researchers say their tests do not establish the independent effect of each wrapper.