AIAny
Icon for item

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

Evaluates and trains multimodal agents to construct interactive 3D open worlds from user queries — provides a large benchmark of assets, seed worlds, and reverse-synthesized queries plus a sandbox RL gym for tool-driven editing and rubric-based verification. Reports that frontier MLLMs perform under 60% and that RL fine-tuning improves precise 3D editing.

Introduction

Why this matters Multimodal agents that can translate a user prompt into a navigable 3D world would enable faster content creation for games, simulation, and embodied AI research — but current evaluations use overly simple queries and lack standardized tooling to measure real editing and interaction abilities. VibeWorlding reframes this gap by defining the "vibe worlding" task and delivering both a benchmark and a training/evaluation gym that require agents to plan, call 3D tools, perform edits, render observations, and verify outcomes over multi-turn interactions.

Key Findings
  • Dataset & benchmark scale: VWE-BENCH contains 2,616 curated 3D assets, 323 human-annotated seed 3D worlds, and 6,828 reverse-synthesized multimodal user queries split into verified (with ground truth) and unverified (rubric-checked) sets — enabling diverse, realistic evaluation scenarios.
  • Evaluation platform: VibeWorlding-Gym unifies asset retrieval, editing tools, and image rendering as MCP-style tools and pairs them with a rubric-based verifier combining physical feasibility and intent-fulfillment checks, making reward signals usable for RL fine-tuning.
  • Empirical result: Contemporary frontier MLLMs (including reported GPT-5.5 and Qwen3.8-Max evaluations) achieve below 60% success rate on vibe worlding tasks, with precise 3D editing identified as the primary bottleneck; RL post-training narrows this gap and enables open-source models (VibeWorlder-8B and VibeWorlder-30B-A3B) to reach competitive or leading Pass@1 performance.
Who it's for and tradeoffs

Great fit if you need a standardized, multi-turn benchmark and training environment for agents that must: interpret multimodal user intent, plan scene layout, invoke 3D editing tools, and verify outcomes using both visual and physical rubrics. The resource is valuable for researchers building embodied or content-creation agents and for teams studying RL-based fine-tuning of MLLMs for tool use. Look elsewhere if your focus is single-frame view synthesis, purely generative 3D modeling without tool-in-the-loop editing, or real-time production pipelines where latency and integration constraints outweigh the benefits of rubric-driven RL training.

Methods & practical notes

The paper emphasizes a joint evaluation+training loop: (1) generate user queries by reverse-synthesis from ground-truth scenes, (2) expose agents to a sandbox of discrete MCP tools (retrieve, edit, render), and (3) use a rubric verifier to produce reward signals for RL. Reproducing the reported RL gains requires the provided benchmark splits and the gym infrastructure; the paper reports that RL improves precise 3D editing more than zero-shot prompting alone.

Information

  • Websitearxiv.org
  • OrganizationsAffiliation: AI Thrust, HKUST(GZ), TEG AIPD, Tencentyning092connect.hkust-gz.edu.cn, [email protected]{jingwenye,wadewdzhang}@tencent.com
  • AuthorsYansong Ning, Jingwen Ye, Zhongkai Wu, Yang Sun, Yiqin Zhu, Xingyi Li, Weidong Zhang, Hao Liu
  • Published date2026/08/15

Categories

More Items

Defines "agentic transactions" and an ACID-style reliability framework for LLM agents that manage long-horizon tasks over persistent environments. Implements an ACID-compliant data agent using exploration–execution–validation cycles, confidence-divergence checks, semantic isolation, and append-only durable workspaces.

Trains a foundation GUI agent using a closed-loop, environment-grounded data stack plus in-context multimodal demonstrations to automate long-horizon desktop workflows. Combines scalable task generation/verification, subtask-level demo guidance, and a 100-task OSWorkerBench benchmark to improve strict success and task progress.

Presents a unified black-box reinforcement learning framework to train and optimize agents running inside complex execution harnesses. Uses sandbox-parallel rollouts, a serving proxy that captures model calls and reconstructs multi-turn trajectories as prefix trees, and adapted GRPO/PPO optimizers to achieve stable, scalable RL across heterogeneous harnesses.