LLM steering planner
My side project this quarter has been building a system that adds steering / planning to LLMs/WAMs/etc. (T, DiT, etc.) in a market/physics/other-ground-truth-constraints aware way. I have a small model outperforming some of the largest frontier models in some games (i.e., adversarial settings) now, but that is not very impressive because "plain old LLMs" are the wrong mathematical object for gameplay, IMO. So far, it's a mix of RL + optimization (e.g., MILP) on the planning side, and not quite realtime enough to play RTSes, but that's only a few weeks away. Have been very impressed at how far the open source ecosystem for this stuff has gotten since the last time I tried something like this.