How to Reliably Gauge LLM Confidence?
a { text-decoration: none; color: #464feb; } tr th, tr td { border: 1px solid #e6e6e6; } tr th { background-color: #f5f5f5; } I’m trying to estimate an LLM’s confidence in its answers in a way that correlates with correctness. Self-reported confidence is often misleading, and raw token probabilities mostly reflect fluency rather than truth. I don’t have grounding options like RAG, human feedback, or online search, so I’m looking for approaches that work in this constraint. What techniques have you found effective—entropy-based signals, calibration (temperature scaling), self-evaluation, or others? Any best practices for making confidence scores actionable?
Doc Intelligence: Custom Extraction model | Confidence score deterioration with new formats/layouts
Hi everyone, This is my first time using custom extraction models on the Document Intelligence service, and I would appreciate your input on an experiment I am conducting. I wanted to investigate how these models' confidence scores behave when documents with significantly different format/layout are introduced (later) in the training phase. I started by training models with documents in the same format (some of worse picture quality and slightly rotated), increasingly adding more samples (a new model was trained every time I added new documents, at increments of 5). After every new model was trained, I checked scores against the same, unseen by the model holdout set that had the same format with those in the training set. After training the final model, with 35 identically formatted documents, I started introducing documents with a significantly different format/layout and retraining (at increments of 10). Confidence scores against the holdout set (unchanged) dropped after doing so, without recovering to previous levels. See graph below showing how confidence scores evolved after every training step (adding new documents at every step). Any insights as to why this has happened?
