WorldRoamBench: An Open-World Benchmark for Long-Horizon Stability of Interactive World Models

2026-06-30

AI Summary

Current evaluations of interactive world models are largely confined to trajectory-level action following, often overlooking critical aspects such as memory retention, physical interaction fidelity, and long-term stability. To address this gap, this work introduces WorldRoamBench—an open-world benchmark that holistically assesses model performance across four dimensions: action execution, visual consistency, physical plausibility, and memory capacity during extended interactive episodes. The benchmark incorporates fine-grained frame-wise action metrics, segment-level visual drift detection, controllability-gated physical evaluation, and a memory testing protocol based on action disentanglement. Leveraging 3D point cloud reconstruction, vision-language reasoning, cross-view modeling, and continuous WASD-based interaction, we evaluate over a dozen state-of-the-art models across more than 600 diverse scenes spanning natural, urban, and indoor environments. Results reveal that no existing model fully satisfies all four criteria, with the best-performing systems achieving only moderate scores, thereby exposing fundamental limitations in current approaches.

Abstract

Despite rapid progress in interactive world models (IWMs), existing benchmarks evaluate action following only at trajectory level and ignore memory and interaction physics. We introduce WorldRoamBench, an open-world benchmark for long-horizon stability across four dimensions, each with tailored innovations: (i) Action: per-frame action metric bypassing cross-model semantic scale disparity and exposing failures hidden by trajectory; (ii) Vision: segment-based drift metric capturing non-monotonic mid-sequence collapse missed by start-vs-end comparisons; (iii) Physics: controllability-gated evaluation over mechanics, optics, and 3D consistency, scoring plausibility under faithful action execution; (iv) Memory: action-decoupled protocol evaluating scene memory via transition-localized 3D point-cloud reconstruction and subject memory via tracking-plus-VLM reasoning. The benchmark comprises 600+ test cases across Nature, Urban, and Indoor scenes in first/third-person views with WASD 10-60s continuous interaction. Evaluating 10+ open/closed-source models reveals none reliably satisfies all dimensions; even the best achieves only moderate scores. Advances on WorldRoamBench are steps toward IWMs that are stable, physically grounded, memory-faithful, and deployable in real-world applications.

Problem

• interactive world models
• long-horizon stability
• memory
• physics
• benchmark

Innovation

• interactive world models
• long-horizon stability
• open-world benchmark
• physics-grounded evaluation
• memory-faithful reasoning