[arXiv]score: 0.39
Double Triangle Annotation Uses Cross-Model Consensus to Scale Historical Document Labeling
May 26, 2026
A two-layer human-in-the-loop framework auto-accepts labels when two independent MLLMs agree and escalates disagreements to human reviewers, reducing annotation cost for historical document extraction under an error-independence assumption.
cs.CL
HOW THIS AFFECTS YOU
●
builderPractical annotation pipeline for structured extraction tasks where LLM hallucination risk is high and ground truth is expensive to obtain.
●
researcherError-independence assumption between architecturally distinct models is a testable and potentially fragile foundation worth stress-testing.