AutoWorldModel-Bench: AI agents improve world models in 91% of runs

World modeling remains an unsettled field, with no single architecture or training objective dominating across environments. AutoWorldModel-Bench, introduced by researchers Marjan Moodi, Xuankang Zhu, Fernando De Mesentier Silva, Harold Chaput, and Mohammad Reza Taesiri, turns this uncertainty into an advantage: instead of measuring AI agents on engineering-to-spec tasks, it evaluates them on open-ended research. The closed-loop benchmark spans 8 game environments, each with ground-truth entity states extracted and fed through a unified structured-state tensor format. This cleanly separates dynamics modeling from perception, letting agents iterate in minutes per run under a fixed compute budget.

In 64 sessions, frontier coding agents—OpenAI's Codex-5.4 and Anthropic's Claude Opus 4.6—successfully improved the provided world-model starter in 63 sessions. Crucially, 91% of winning edits were substantive research-style modifications: new objectives, state representations, rollout procedures, or architectural changes, rather than simple hyperparameter tweaks. This suggests these agents can handle long-horizon, undirected problem-solving, a capability most existing benchmarks ignore. AutoWorldModel-Bench offers the first testbed where autonomous coding agents can be graded on genuine research progress, potentially reshaping how AI capabilities for scientific discovery are measured.