Evaluation shows claim-level verification and isotonic calibration reduce error in language models, indicating improved confidence estimates.
Key Points
To address the overconfidence and unreliability of large language models by developing and evaluating a framework that decomposes answers into atomic claims for verification and calibration.
Designed CLAIM-CAL, a framework that breaks generated responses into atomic factual claims, tests them with multiple verification probes, calculates a risk score, and applies post-hoc isotonic calibration.
Evaluated the pipeline on TruthfulQA generation using 200 examples (60 calibration, 140 held-out test), benchmarking against direct answering, verbal confidence, self-consistency, and simple self-verification.
Isotonic calibration reduced the Expected Calibration Error (ECE) of raw CLAIM-CAL from 0.212 to 0.038 [95% CI: 0.015, 0.106], while maintaining an accuracy of 0.757.
In calibrated baseline comparisons, calibrated CLAIM-CAL attained an ECE of 0.038, a Brier Score of 0.146, and a selective accuracy of 0.824 at a 0.7 confidence threshold with 0.893 coverage.
A second-pass judge prompt-robustness check across 50 evaluations demonstrated 96.0% agreement and a Cohen's kappa of 0.896.