WebAnything — reconstructing reusable web environments

TL;DR

WebAnything reconstructs executable, resettable website environments from real URLs and separates construction from verification. The authors report transfer gains after training a web agent on trajectories collected in these environments. (source)

Authors & Org

Tianci Xie, Zikun Cai, Chaoyun Zhang, Pinxue Guo, Bingxuan Li, Rui Zheng and Hang Yan. Affiliations are not established in the captured abstract record. First submitted 2026-10-27. (source)

Method

An Explorer-Builder collects browser-visible structures, resources and interaction outcomes. An independent verifier checks frozen candidate environments and guides local repairs or targeted recapture from the source. The authors describe no per-site manual implementation or repair after initial setup and authorization; this does not remove those initial requirements. (source)

Results

The collection contains 244 website artifacts across 14 application domains and 53 categories, supporting 12,984 executed and accepted tasks. Against WebHarbor, the authors report higher aggregate scores on three fidelity metrics, with 12 wins, two ties and no losses in a blind audit. Fine-tuning Qwen3.5-9B on 5,856 trajectories improves REAL success by 9.9 percentage points, Odysseys-Easy Rubric Avg and Task Avg by 6.25 and 6.32 percentage points, and WebArena-Verified Overall success by 13.33 percentage points, under a fixed harness. (source)

Significance

For Agents (LLM Agents) and Eval Harness Configuration, the contribution is a reusable environment-production loop with a separate verifier. Reported transfer is stronger evidence than visual resemblance alone, while remaining specific to the evaluated model and harness. (source)

Open Questions

The abstract does not establish reconstruction cost, robustness to website updates, coverage of unobserved interactions or independence from evaluation contamination. Full-paper methods, code and independent reproduction were not reviewed. (source)

Cite

• arXiv:2610.37786.
• Captured abstract.