Parallel WebBench - Element-wise F1 (WebExplorer-Style Scaffold): leaderboard
Parallel web exploration (Parallel WebBench, Sogani et al., 2026): multi-field questions whose answers must be gathered from several subpages of a live website and combined into one structured answer, 150 held-out tasks, 16k-token context and 16 interaction rounds. Run as a WebExplorer-style agent with search and browse tools; the score is element-wise F1 of the answer, judged by GPT-4.1-mini against the gold answer, in percent; higher is better.
Metric: Element-wise F1 (%; GPT-4.1-mini judge, per answer element) on the 150 held-out tasks, 16k input and 16 interaction rounds, WebExplorer-Style Scaffold. Source: arxiv.org. Saturation forecast: Around February 2027. 6 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-4.1 Mini | 39.49 | #775 |
| 2 | Qwen 2.5 7B Instruct | 21.08 | #2583 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=parallel-webbench-element-wise-f1-webexplorer-style-scaffold · How It Works · Data refreshed daily, snapshot 2026-11-12.