K-BrowseComp - Calibration Error: leaderboard
Calibration on K-BrowseComp (2026), a Korean web-browsing agent benchmark: the expected calibration error in percent between the confidence an agent states and its accuracy on the 300 verified questions, over five equal-width bins. Lower is better.
Metric: Expected calibration error (%) of the stated confidence on K-BrowseComp-Verified (five equal-width bins, weighted gap between mean confidence and accuracy); lower is better. Source: arxiv.org. Saturation forecast: Around February 2027. 10 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | DeepSeek V4 Pro | 17.72 | #188 |
| 2 | Gemma 4 31B (IT) | 23.66 | #342 |
| 3 | K-EXAONE | 24.09 | #708 |
| 4 | GLM-5.1 | 27.07 | #179 |
| 5 | GPT-5.5 | 31.86 | #49 |
| 6 | GPT-5.4 Mini | 37.88 | #389 |
| 7 | Qwen 3.6 35B A3B | 47.89 | #462 |
| 8 | A.X 4.0 | 47.89 | #863 |
| 9 | Gemini 3.1 Flash Lite | 56.55 | #364 |
| 10 | HyperCLOVA X SEED Think (32B) | 77.37 | #1437 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=k-browsecomp-calibration-error · How It Works · Data refreshed daily, snapshot 2026-11-13.