VibeWorlding: Can Multimodal Agents Construct 3D Open Worlds End-to-End?

Hugging Face Daily Papers Papers

Summary

This paper proposes VibeWorlding, a framework for benchmarking and training multimodal agents to construct 3D open worlds from user queries, showing that reinforcement learning improves open-source models to compete with closed-source frontiers.

Constructing an interactive 3D open world from a user query is important. However, existing methods are primarily evaluated on idealized, simple queries, making it difficult to systematically analyze and compare how multimodal agents understand user intent, use 3D tools, and reason over textual and visual 3D world information. To this end, we propose VibeWorlding, a unified framework for benchmarking and training vibe worlding agents: a multimodal agent that can autonomously infer user intent, plan scene layout, invoke 3D tools, and reflect on the multimodal feedback in a multi-turn agent-environment interaction process. To achieve this, we first build VWE-BENCH, a benchmark of 2,616 high-quality 3D assets, 323 human-annotated seed 3D worlds, and 6,828 reverse-synthesized multimodal user queries, split into verified queries with ground-truth and unverified queries with carefully designed rubrics. Moreover, we develop VibeWorlding-Gym, a joint multimodal RL post-training framework that integrates (1) a sandbox environment unifying asset retrieval, editing, and image rendering as MCP tools, and (2) a rubric-based verifier that combines physical feasibility and intent fulfillment verification, supporting both fair model evaluation and scalable multimodal RL reward service. Our experiments show that current frontier MLLMs are far from solving the vibe worlding agent task, with even GPT-5.5 and Qwen3.8-Max reaching below 60% success rate, and trace the bottleneck to precise 3D world editing. We further find that RL training can ease this weakness and enable open-source MLLMs to even surpass closed-source frontiers: our VibeWorlder-8B is comparable to frontier MLLMs, while our flagship VibeWorlder-30B-A3B attains the best overall Pass@1 among all evaluated models.
Original Article
View Cached Full Text

Cached at: 08/18/26, 03:53 AM

Paper page - VibeWorlding: Can Multimodal Agents Construct 3D Open Worlds End-to-End?

Source: https://huggingface.co/papers/2608.15265

Abstract

A unified framework benchmarks and trains multimodal agents that infer intent, plan 3D scenes, invoke tools, and reflect on feedback, revealing that reinforcement learning improves open-source models beyond closed-source frontiers.

Constructing an interactive 3D open world from a user query is important. However, existing methods are primarily evaluated on idealized, simple queries, making it difficult to systematically analyze and compare howmultimodal agentsunderstand user intent, use 3D tools, and reason over textual and visual 3D world information. To this end, we proposeVibeWorlding, a unified framework for benchmarking and training vibe worlding agents: a multimodal agent that can autonomously infer user intent, plan scene layout, invoke 3D tools, and reflect on the multimodal feedback in a multi-turn agent-environment interaction process. To achieve this, we first buildVWE-BENCH, a benchmark of 2,616 high-quality 3D assets, 323 human-annotated seed 3D worlds, and 6,828 reverse-synthesized multimodal user queries, split into verified queries with ground-truth and unverified queries with carefully designed rubrics. Moreover, we developVibeWorlding-Gym, a jointmultimodal RLpost-training framework that integrates (1) a sandbox environment unifying asset retrieval, editing, and image rendering asMCP tools, and (2) arubric-based verifierthat combines physical feasibility and intent fulfillment verification, supporting both fair model evaluation and scalablemultimodal RLreward service. Our experiments show that current frontierMLLMsare far from solving the vibe worlding agent task, with even GPT-5.5 and Qwen3.8-Max reaching below 60% success rate, and trace the bottleneck to precise 3D world editing. We further find that RL training can ease this weakness and enable open-sourceMLLMsto even surpass closed-source frontiers: ourVibeWorlder-8B is comparable to frontierMLLMs, while our flagshipVibeWorlder-30B-A3B attains the best overall Pass@1 among all evaluated models.

View arXiv pageView PDFProject pageGitHub24Add to collection

Models citing this paper0

No model linking this paper

Cite arxiv.org/abs/2608.15265 in a model README.md to link it from this page.

Datasets citing this paper0

No dataset linking this paper

Cite arxiv.org/abs/2608.15265 in a dataset README.md to link it from this page.

Spaces citing this paper0

No Space linking this paper

Cite arxiv.org/abs/2608.15265 in a Space README.md to link it from this page.

Collections including this paper1

Similar Articles

Infinite Worlds with Versatile Interactions

Hugging Face Daily Papers

This paper introduces LingBot-World 2.0, an advanced world modeling system with unbounded interaction horizons, real-time 720p 60fps video generation, diverse interactive elements (e.g., attacking, spell-casting), and novel multi-agent behavior control via pilot and director agents, along with a shared multi-player interface.

MultiWorld: Scalable Multi-Agent Multi-View Video World Models

Hugging Face Daily Papers

MultiWorld is a unified framework for multi-agent multi-view video world modeling that achieves accurate control of multiple agents while maintaining multi-view consistency through a Multi-Agent Condition Module and Global State Encoder.

MineExplorer: Evaluating Open-World Exploration of MLLM Agents in Minecraft

Hugging Face Daily Papers

The MineExplorer benchmark evaluates multimodal large language model agents' open-world exploration abilities in Minecraft using atomic and multi-hop tasks designed through multi-agent synthesis. Experiments show that open-world exploration remains challenging, with strong models degrading sharply over longer trajectories.