CORDIAL calibrates ordinal LLM outputs using minimal labels
September 25, 2026
CORDIAL corrects noisy LLM distributions over ordered scales using a five-parameter calibration channel. In 76 of 80 tests, it achieved the lowest log loss among nine calibrators using only 5 to 100 labels, matching strong baselines with significantly less data.
HOW THIS AFFECTS YOU
●
builderYou can use this to calibrate LLM-based rating or ranking systems with very few human-labeled examples.
●
researcherThe method provides a way to preserve first-order stochastic order during calibration with minimal supervision.