Game Coach Gets Strategy Conditions, Not Facts

🔥 What happened Researchers at Tencent built TactiNavi, the serving system behind an LLM game coach in a commercial open-world title. Instead of checking atomic facts, it compiles versioned strategy packages with explicit prerequisites and evaluates them online against structured player state. 💡 Why it matters Across 1,122 logged coaching turns, correctness jumped from 57.9% to 82.0%, and strategy misapplication under changed player state dropped from 18.1% to 10.4%. The kicker: an unchanged-system rerun recovers 38.4% of failing turns, and only 1 of 15 shipped revisions passes release checks. Most apparent fixes are generation noise. ⚡ Our take If your agent release path lacks a null model, you're tuning noise. TactiNavi treats applicability as a first-class serving object — engineers should steal that idea before shipping another “fix.”