Benchmark RadarConfBench
ConfBench is a calibration-specific benchmark for key information extraction from documents. It applies 20 degradation pipelines to create 1,346 variants and over 70K entity-level evaluations, spanning the accuracy spectrum for evaluating confidence estimates of VLMs.
- Released
- 2026-08-03
- Readiness
- Paper only
- Primary field
- General AI
Why it matters
Document processing requires trustworthy confidence scores for routing automation vs. human review. ConfBench enables systematic study of confidence estimators and calibration methods, addressing the lack of calibration-focused benchmarks.
Motivation
Intelligent document processing (IDP) with vision-language models (VLMs) hinges on confidence scores trustworthy enough to route extractions between automation and human review.
Primary resources
Benchmark Radar records only publicly supported details and links back to primary sources for verification.