$ cat wiki/papers/2026/2610.30181-opends-world.md
OpenDS-World
TL;DR
OpenDS-World evaluates live-web research that must connect evidence into one factual answer. The strongest tested model reaches 39.0% accuracy on 200 tasks. (source)
Authors & Org
Xiaomin Yu, Hongliang Chen, Zhe Huang, Shilin Yan, Shao-Rong Su, Xuqiao Ma, Feiyang Wang, Jiayu Chen, Jiwei Fang, Yu Yan, Yini Huang, Chen Chen, Ruoxi Cheng, Yuzhi He, Yuhui Zhang, Xiaoxing Hu, Chen Liu, Jianbo Lin, Yijiang Li, Xiaobin Hu, Menglong Li, Zhuoran Lu, Weishi Wang, Zhongqi Yue, Chengwei Qin. The abstract page does not establish affiliations. First submitted 2026-10-19. (source)
Method
OpenDS combines web search, browser interaction and evidence processing across modalities, file formats and access mechanisms. Independent human checks establish evidence validity and answer uniqueness. The authors evaluate 12 frontier models in a unified tool environment and execution protocol. (source)
Results
The best model scores 39.0% accuracy; human evaluation across 1,000 independent attempts scores 21.1%. The abstract does not state matched time budgets or participant expertise, so this is not evidence of general superiority to human researchers. Analyses identify differences in tool coordination, termination and failure recovery. (source)
Significance
For Agents (LLM Agents), retrieving documents is only part of research: access, evidence integration and deciding when to stop are tested together. Eval Harness Configuration should retain the execution protocol when comparing results. (source)
Open Questions
The abstract does not name the winning configuration or specify costs, human time limits, or reproducibility under changing web content. Full paper and code have not been reviewed here. (source)
Cite
arXiv:2610.30181 · Immutable abstract capture. (source)