JEDI World Model Reverses Performance Asymmetry

An arXiv paper (v3, Feb 24, 2026) by Jing Yu Lim et al. finds a pronounced Performance Asymmetry in model-based RL on the Atari100k benchmark, with a 21× gap between Human-Optimal and Agent-Optimal task subsets. The authors introduce Sym-HNS and propose JEDI, a latent end-to-end Joint Embedding DIffusion world model that achieves SOTA on Sym-HNS, improves Human-Optimal and Breakout performance, and boosts computational efficiency.