ConfBench - Expected Calibration Error: leaderboard
ConfBench (Amazon, 2026) degrades 75 verified invoices with 20 controlled augmentation pipelines into 1,346 variants and asks a vision-language model to extract entities and state a confidence for each, here as verbalized top-4 guesses from the OCR text plus the image. Score is the expected calibration error between stated confidence and observed accuracy, on a 0-1 scale; lower is better.
Metric: Expected calibration error of the verbalized confidence (0-1, lower is better; adaptive 5-bin quantile binning; verbalized top-4 guesses with probabilities in one call, OCR text with Amazon Textract confidences plus the document image, over the entity fields of 1,346 degraded variants of 75 RealKIE-FCC-Verified invoices). Source: arxiv.org. Saturation forecast: Estimated already saturated. 7 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Opus 4.6 | 0.05 | #102 |
| 2 | Claude Sonnet 4.5 | 0.13 | #250 |
| 3 | Claude Haiku 4.5 | 0.17 | #568 |
| 4 | Kimi K2.5 | 0.2 | #256 |
| 5 | Qwen 3.6 27B | 0.22 | #341 |
| 6 | Qwen 3 VL 235B A22B | 0.25 | #479 |
| 7 | Gemma 3 12B | 0.31 | #2000 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=confbench-expected-calibration-error · How It Works · Data refreshed daily, snapshot 2026-11-13.