Ko-WideSearch - Item F1: leaderboard
Evaluates Korean broad web research with structured table outputs on 228 tables in a single-pass protocol. Item, column, row and complete-table success are distinct published metrics; parse rate is not an extra board. This series reports Item F1 using F1 (%). Source: the original paper result table.
Metric: F1 (%). Source: arxiv.org. Saturation forecast: Estimated already saturated. 20 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Opus 4.7 | 94.6 | #51 |
| 2 | Claude Opus 4.8 | 94.1 | #39 |
| 3 | GPT-5.5 | 92.8 | #33 |
| 4 | Claude Opus 4.6 | 92 | #67 |
| 5 | Claude Sonnet 4.6 | 90.2 | #94 |
| 6 | GPT-5.4 | 89 | #83 |
| 7 | Gemini 3.1 Flash Lite | 89 | #200 |
| 8 | Gemini 3.1 Pro (Preview) | 88.2 | #64 |
| 9 | GPT-5.4 Mini | 82.3 | #228 |
| 10 | DeepSeek V4 Pro | 80.4 | #112 |
| 11 | Gemini 3.5 Flash | 79.9 | #66 |
| 12 | Gemma 4 31B | 76.4 | #212 |
| 13 | A.X 4.0 | 71.7 | #436 |
| 14 | Claude Haiku 4.5 | 66.7 | #287 |
| 15 | GPT-5.4 Nano | 66.6 | #359 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=ko-widesearch-item-f1 · How It Works · Data refreshed daily, snapshot 2026-11-05.