This content explores three post-hoc methods—Platt Scaling, Isotonic Regression, and Temperature Scaling—designed to enhance the calibration of language models. These techniques aim to reduce the disparity between a model's predicted confidence and its actual accuracy.
The size of the human-labeled calibration set for validating an LLM-as-judge depends on label balance. Fifty stratified traces suffice for balanced binary criteria, but 200 or more are mandatory for rare-but-expensive categories like safety violations, as kappa's variance is dominated by minority-class examples.
This guide treats Nvidia's open-source Ising quantum AI models as production services, focusing on their deployment, orchestration, guardrails, and governance within existing AI security frameworks. It highlights the critical importance of calibration for the real-world performance of quantum-inspired Ising solvers, as mis-tuned systems can lead to significant production failures.
Large language models are often overconfident, expressing high certainty even when incorrect. This paper introduces SECL, a test-time training pipeline that exploits a self-supervised signal to improve calibration without requiring labeled data or human supervision.
This research addresses the lack of diversity in LLM outputs, attributing it to how models allocate probability mass across valid and invalid continuations during decoding. It introduces a validity-diversity framework that decomposes the problem into two complementary forms of miscalibration: order calibration and shape calibration.
This paper formalizes calibration for probabilistic label ranking, introducing a hierarchy of notions for full, sub-ranking, and top-k calibration. Empirically, popular label ranking models are often poorly calibrated, with implications for RLHF reward models.