World-in-World: World Models in a Closed-Loop World

TL;DR

World-in-World addresses the problem of evaluating world models for embodied agents in closed-loop interactions, where predictive perception must translate into decision-making and action. It introduces a unified closed-loop planning framework with proposal–simulation–revision loops and a standardized action API that accommodates diverse world models, evaluated across four embodied tasks (AR, ImageNav, A-EQA, RLBench manipulation). A key contribution is the data-scaling post-training protocol that adapts pretrained video generators to target domains, revealing a data-scaling law and showing that action-conditioned fine-tuning can outperform simply upgrading the pretrained generator. The experiments demonstrate that high visual quality alone does not guarantee task success; scaling downstream data and inference-time compute substantially improves performance, highlighting the practical utility of world models for embodied decision-making and guiding future WM research toward controllability, adaptation, and robust planning. The work provides a valuable benchmark and framework for measuring embodied utility, with implications for deploying predictive world models in real-world autonomous systems and simulations alike.

Abstract

Generative world models (WMs) can now simulate worlds with striking visual realism, which naturally raises the question of whether they can endow embodied agents with predictive perception for decision making. Progress on this question has been limited by fragmented evaluation: most existing benchmarks adopt open-loop protocols that emphasize visual quality in isolation, leaving the core issue of embodied utility unresolved, i.e., do WMs actually help agents succeed at embodied tasks? To address this gap, we introduce World-in-World, the first open platform that benchmarks WMs in a closed-loop world that mirrors real agent-environment interactions. World-in-World provides a unified online planning strategy and a standardized action API, enabling heterogeneous WMs for decision making. We curate four closed-loop environments that rigorously evaluate diverse WMs, prioritize task success as the primary metric, and move beyond the common focus on visual quality; we also present the first data scaling law for world models in embodied settings. Our study uncovers three surprises: (1) visual quality alone does not guarantee task success, controllability matters more; (2) scaling post-training with action-observation data is more effective than upgrading the pretrained video generators; and (3) allocating more inference-time compute allows WMs to substantially improve closed-loop performance.

Figures (19)

• Figure 1: We introduce the first open benchmark to evaluate world models by closed-loop task success, analyze the link between task success and visual quality, and investigate scaling laws.
• Figure 2: Task success rate vs. generation quality. $\dagger$: post-trained with extra data. We defend that world models live and die by their closed-loop success, not flawless generated visuals.
• Figure 3: Closed-loop online planning in World-in-World: At time step $t$, the agent receives the world state, represented by observation $\mathbf{o}_{t}$, and invokes a proposal policy $\pi_{\text{proposal}}$ (❶) to produce a total of $M$ candidate action plans. The unified action API (❷) transforms each plan into the control inputs required by the world model. The world model (❸) then predicts the corresponding future states as observations $\hat{\mathbf{O}}_{t}$. The revision policy $\pi_{\text{revision}}$ (❹) evaluates all rollouts and commits to the best, yielding decision $\mathbf{D}^{\star}_{t}$. This decision is applied in the environment, closing the interaction loop.
• Figure 4: Top-left: Active Recognition (AR), the agent needs to identify a designated target under occlusions or extreme viewpoints while minimizing navigation cost. Top-right: Image-Goal Navigation (ImageNav), the agent reaches the viewpoint matching a goal image, emphasizing success rate and path efficiency. Bottom-left: Active Embodied Question Answering (A-EQA), the agent answers an open-ended question after active exploration. Bottom-right: Robotic Manipulation, the agent needs to control a robotic arm to complete tasks such as grasping and placement to specified targets.
• Figure 5: (a) SR vs. generation quality in AR; generation quality is scored as the average of an aesthetic predictor aesthetic_predictor_v2.5 and an image-quality predictor Ke2021MUSIQMI, both trained to match human preferences. (b) SR vs. controllability in AR; controllability is quantified as $1-\mathrm{LPIPS}$ between ground-truth and predicted observations.