AutoWorldModel-Bench: A State-Centric Benchmark for Automated World-Model Research

2026-07-20

AI Summary

This study addresses the absence of established research paradigms for world models and the inadequacy of existing benchmarks in evaluating open-ended scientific inquiry capabilities. We propose the first automated closed-loop benchmark specifically designed for world model research. By employing unified structured state representations to decouple perception from dynamics modeling, our framework enables coding agents to autonomously iterate and optimize models under fixed computational budgets. Experimental results demonstrate that 94% of sessions achieved performance improvements, with half yielding significant gains; notably, 91% of superior edits constituted substantive refinements to model architectures or training protocols. This work effectively validates agents' capacity for open-ended research and establishes a standardized evaluation framework for advancing world model development.

Abstract

World modeling is an unsettled field: architectures, training objectives, and state representations interact in complex ways, and no single recipe dominates across environments. This makes it an ideal testbed for AI coding agents acting as autonomous researchers--a setting in which the improvement direction is not specified in advance, unlike the engineering-to-spec tasks that dominate current agent benchmarks. We 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--ground-truth entity state extracted from each game and consumed through a shared tensor format--which isolates dynamics modeling from perception and enables minutes-per-run iteration. Across 64 sessions, Codex-5.4 and Claude Opus 4.6 improve their starter on 63; in 91% of sessions the winning edit is a non-trivial research-style modification--a new objective, representation, rollout procedure, or architectural change--rather than a hyperparameter tweak. Our benchmark offers a setting in which frontier coding agents can be evaluated on open-ended research rather than engineering-to-spec problems.

Problem

• World Model
• AI Coding Agents
• Benchmark
• Autonomous Research

Innovation

• State-Centric Benchmark
• Automated World-Model Research
• Structured-State Representation
• Closed-Loop Evaluation
• AI Coding Agents