From confidence to actionThresholds searched on labelled answers: when an LLM-backed app should act, ask or hand over14 min read·liminma.hashnode.dev2d ago
Does confidence separate right from wrong?Accuracy-coverage curves and AUROC on an LLM's typed decisions14 min read·liminma.hashnode.dev5d ago
When one knob isn't enough: isotonic regressionFitting a flexible calibration curve to an LLM's confidence and counting the cost13 min read·liminma.hashnode.dev1Nov 5
Fixing the confidence: temperature scalingFitting one number per step on 770 labelled answers to make an LLM's confidence more honest12 min read·liminma.hashnode.dev26KAAOct 29
Is the confidence honest?Building reliability diagrams and computing ECE from scratch to check an LLM's confidence10 min read·liminma.hashnode.devOct 24
