DRILL paper proposes training world models for policy utility, not pixel prediction

Researchers led by Zhaolu Kang released a paper called DRILL that trains video world models to boost policy performance rather than to predict the next frame accurately. The method, presented at NeurIPS 2026, includes two versions: DRILL‑VWI, which reweights prediction errors, and DRILL‑IMG, a meta‑gradient approach. Tests on five manipulation simulators showed DRILL‑IMG improved success rates by 13.1 points over the previous WMPO baseline. The paper argues that lower pixel error does not guarantee better policies.