Uncertainty Quantification in LLMs

LLMs produce a probability distribution over possible next tokens at every step. That distribution encodes how confident the model is. But a confident-sounding answer in English is not the same as the model's internal probability — and the gap between them is where uncertainty quantification lives.

Three kinds of uncertainty

• Aleatoric: noise inherent in the data (different annotators would label differently)
• Epistemic: uncertainty from the model not having seen enough
• Model: uncertainty from choice of architecture or training

Signals you can actually read

Why it matters in practice

1. 1Let low-confidence answers trigger a tool call or human review
2. 2Abstain from answering when uncertainty is too high
3. 3Surface uncertainty in the UI so users can weigh it
4. 4Track calibration over time as a quality metric

“A responsible model should not just give you an answer. It should tell you how much to trust it.”

The big idea: confidence without calibration is noise. Quantifying uncertainty turns an LLM from a slot machine into a sensor.