ConfBench - Error Capture (ECARB@30): 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 how many more errors a reviewer finds by checking the least-confident 30 percent of predictions than by checking a random 30 percent; 1 is no lift, 3.33 the ceiling, and higher is better.
Metric: Error capture at a 30 percent review budget (ECARB; errors found by reviewing the least-confident 30 percent of predictions divided by the errors a random 30 percent review finds, so 1 is no lift and 3.33 the ceiling; 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: Around February 2027. 7 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Opus 4.6 | 2.43 | #102 |
| 2 | Claude Sonnet 4.5 | 2.24 | #250 |
| 3 | Claude Haiku 4.5 | 1.96 | #568 |
| 4 | Qwen 3.6 27B | 1.93 | #341 |
| 5 | Kimi K2.5 | 1.68 | #256 |
| 6 | Qwen 3 VL 235B A22B | 1.53 | #479 |
| 7 | Gemma 3 12B | 1.39 | #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-error-capture-ecarb-30 · How It Works · Data refreshed daily, snapshot 2026-11-13.