AutoWorldModel-Bench: New Benchmark for Autonomous AI World-Model Research

A new benchmark called AutoWorldModel-Bench has been developed by researchers to assess AI coding agents functioning as independent researchers in world modeling. This benchmark is outlined in a paper available on arXiv (arXiv:2608.11216). World modeling is characterized as a complex and evolving area, where various architectures, training goals, and state representations interact without a clear dominant approach. This complexity makes it suitable for AI agents to autonomously enhance their models without predefined improvement paths, unlike conventional engineering tasks. The benchmark includes a foundational world model and a set compute budget, pushing advanced coding agents to refine the model independently. It encompasses eight gaming environments unified under a structured state representation, with true entity states extracted and processed in a common tensor format. This structure separates dynamics modeling from perception, allowing for rapid iterations. In 64 sessions, Codex-5.4 and Claude Opus 4.6 successfully improved their initial models in 63 instances, achieving a 91% success rate. The benchmark's goal is to foster automated AI research, potentially speeding advancements in world modeling and related domains.