An AI framework can now map the gap between how an amateur plays StarCraft II and how a champion would — then chart a path between the two.
Researchers introduced Latent Maps of Performance, a system built on a Guided Variational Autoencoder trained on 23,305 professional StarCraft II tournament replays. The model learns a compressed representation of expert play, then uses counterfactual generation to trace improvement trajectories from a losing player's profile toward winning configurations. Four path-finding strategies were tested — linear interpolation, iterative optimal transport, density-regularized gradient ascent, and neural flow matching — each designed to keep suggested moves grounded in real expert behavior rather than drifting into nonsense. Feedback can be extracted at different levels of granularity to match where a player is in their development.
Chess and Go players have had AI coaching tools for years; StarCraft II, a faster and more complex real-time strategy game, has not had a principled equivalent. This work borrows from sports science methodology — the same championship model framework used to analyze elite athletic performance — and applies it to a domain where human reaction speed and strategic depth interact in ways that make "just watch the pro replay" advice largely useless. The approach could generalize to other real-time strategy games that have similarly rich replay datasets.
The authors are candid that a trade-off exists between the four traversal methods and that no single strategy wins outright — which is a more honest conclusion than most AI papers manage.