World modeling remains an unsettled field in AI, with architectures, training objectives, and state representations interacting in complex ways and no single recipe dominating across environments. This complexity makes it an ideal testbed for AI coding agents acting as autonomous researchers, where the improvement direction is not specified in advance—unlike the engineering-to-spec tasks that dominate current agent benchmarks.
To address this, researchers at Hugging Face introduce AutoWorldModel-Bench, a closed-loop benchmark in which frontier coding agents autonomously improve a provided world-model starter under a fixed compute budget. The benchmark spans eight game environments under a unified structured-state representation, where ground-truth entity state is extracted from each game and consumed through a shared tensor format. This design isolates dynamics modeling from perception and enables minutes-per-run iteration.
In experiments across 64 sessions, two frontier models—Codex-5.4 and Claude Opus 4.6—improved their starter model in 63 sessions. Notably, in 91% of sessions, the winning edit was a non-trivial research-style modification, such as a new objective, representation, rollout procedure, or architectural change, rather than a simple hyperparameter tweak.
The benchmark offers a setting in which frontier coding agents can be evaluated on open-ended research rather than engineering-to-spec problems, potentially guiding future development of autonomous AI research capabilities.