Parallel WebBench - Element-wise F1 (WebWalker-Style Scaffold): leaderboard

Metric: Element-wise F1 (%; GPT-4.1-mini judge, per answer element) on the 150 held-out tasks, 16k input and 16 interaction rounds, WebWalker-Style Scaffold. Source: arxiv.org. Saturation forecast: Around January 2027. 3 models tracked.

Top models

#ModelScore
1GPT-4.1 Mini33.36
2Qwen 2.5 7B Instruct12.09

Interactive version: theaggregate.ai/benchmark?slug=parallel-webbench-element-wise-f1-webwalker-style-scaffold · How It Works · Data refreshed daily, snapshot 2026-09-26.