Omni2Web - Instruction Recovery Target F1: leaderboard

Metric: Target F1 (%; Track B component: recovered edit targets matched one-to-one to the reference targets; 918 bilingual screen recordings of weakly referential webpage edit requests; the model rewrites the recording into an ordered list of explicit target and action instructions without the source HTML; one-to-one matching against reference instructions, semantic equivalence judged by Gemini 3.1 Pro; unusable outputs score zero). Source: arxiv.org. Saturation forecast: Around 2029. 17 models tracked.

Top models

#ModelScore
1Qwen 3.5 Omni Plus49.99
2Claude Opus 548.06
3Seed 2.0 Lite47.91
4GPT-5.6 Sol47.51
5Gemini 3.5 Flash47.04
6Gemini 3.1 Pro (Preview)42.19
7Kimi K338.64
8Qwen 3.8 Max34.84
9Grok 4.634.6
10MiMo-V2.533.95
11Seed 2.1 Pro33.21
12MiniMax-M321.25
13Qwen 3.8 27B19.82
14Muse Spark 1.118.45
15MiniCPM-o-4.517.87

Interactive version: theaggregate.ai/benchmark?slug=omni2web-instruction-recovery-target-f1 · How It Works · Data refreshed daily, snapshot 2026-09-26.