Create and explore imaginative 3D worlds with our new product Dreamfield (opens in a new tab)

Winning on the dire side of the MOBA

Determine whether a dream-trained policy can win games while controlling the dire side of the ten-hero PufferLib MOBA.

Background

The paper evaluates policies against the shipped PufferNet opponent and reports substantial success only when the learned policy controls the radiant side. When controlling the dire side, neither the learned policy nor the shipped opponent wins in the reported experiments. The authors therefore identify performance on the dire side as unresolved, leaving open whether the asymmetry can be overcome by a different policy, training procedure, world model, or evaluation strategy.

References

The dire side is unsolved.

— Learning in Dreams, Winning in Reality: A Continuous Dyna Loop for a Ten-Hero MOBA  (2610.08033 - Kieto, 6 Oct 2026) in Section Limitations