Search-to-World

Search-to-World

Evaluation of 3D World Delivery
from User Request through Web Search

Zixiao Gu1,2,3,4 Yabo Chen1,† Xunzhi Xiang1,5 Yu He3,4,6 Haibin Huang1 Chi Zhang1 Yunbo Wang2,* Xuelong Li1,*
1Institute of Artificial Intelligence, China Telecom (TeleAI) 2Shanghai Jiao Tong University 3Shanghai Innovation Institution 4ActiMind 5Nanjing University 6Fudan University

†Project Leader*Corresponding Authors

Paper · arXiv

From a request to a delivered world.

Retrieval is the opportunity. Delivery is the outcome.

Search-to-World teaser: a user request leads to web evidence, WorldSearcher, and an evaluated 3D world.
Search-to-World connects user requests, live-web visual evidence, and usable 3D worlds.
Demo A complete walkthrough, from web search to delivery and recovery.

Task overview

Observe relevant content. Deliver a request-aligned, perceptually acceptable world.
Search-to-World measures these as two distinct outcomes.

Task overview: a user request, web retrieval, candidate world, request alignment and perceptual quality assessment, and delivery result.
ORR — Observed Retrieval Rate   ·   WDR — World Delivery Rate

Abstract

Agentic systems can understand user requests, search the live web, and use external tools to complete complex tasks, yet whether they can carry a user request through relevant content web retrieval to a usable 3D world has not been systematically evaluated. Both an established end-to-end pipeline and benchmark for evaluating it are missing. The core problem is bridging the gap from retrieval to delivery: relevant web-sourced visual content must be successfully turned into a request-aligned and perceptually acceptable 3D world.

We address these gaps separately. First, we introduce Search-to-World, an evaluation task that defines an end-to-end pipeline from a user request through web-sourced visual content retrieval to the delivery of a usable 3D world. Observed Retrieval Rate (ORR) and World Delivery Rate (WDR) are proposed as two metrics to distinguish observing relevant web-sourced visual content from successfully delivering a 3D world. Second, we present WorldSearcher, a reuse-then-reconstruction harness system that connects existing search agent systems to 3D world delivery. It first seeks available 3D worlds for reuse, and then falls back to video-based reconstruction when no reusable worlds are retrieved.

We also design a structured recovery control in WorldSearcher to revise temporal grounding, replace source videos, or reformulate queries for failure recovery. We benchmark several widely used models on the Search-to-World task using WorldSearcher and conduct an ablation study of supervised fine-tuning (SFT) for the core models used by recovery subagents, thereby validating the recovery control of WorldSearcher. Our results reveal that Search-to-World delivery varies with the model used by the agentic system and that observing relevant web-sourced visual content alone does not guarantee successful delivery. More importantly, joint training of the recovery agents improves both delivery success and action efficiency.

Overall, Search-to-World makes the capability of 3D world delivery by an agentic system measurable, while WorldSearcher provides an effective harness system with complementary recovery components.

WorldSearcher

A reuse-first harness with video reconstruction and structured recovery.

WorldSearcher method: search for reusable 3D resources, fall back to video reconstruction, assess world alignment and quality, and apply structured recovery.
Reuse and reconstruction share the same world-assessment gate.

Reuse first. Retrieve an existing 3D world aligned with the request.

Reconstruct when needed. Ground a source video in time and reconstruct a candidate world.

Recover with structure. Revise the time interval, replace the video, or reformulate the query.

BibTeX

@article{gu2026searchtoworld,
  title   = {Search-to-World: Evaluation of 3D World Delivery
             from User Request through Web Search},
  author  = {Gu, Zixiao and Chen, Yabo and Xiang, Xunzhi and
             He, Yu and Huang, Haibin and Zhang, Chi and
             Wang, Yunbo and Li, Xuelong},
  journal = {arXiv preprint arXiv:2609.07605},
  year    = {2026}
}