Full talk · edited transcript

When Models Should Stay Silent

Michael Diskin23 May 2026

Michael Diskin, HSE University.

The talk is a position-style overview of the field now usually called hallucination detection, which academia previously called uncertainty quantification (UQ). The goal: when a model answers, measure whether the model "thinks" the answer is correct, or otherwise estimate the probability that the answer is correct. The talk covers why plain confidence does not work for LLMs, which signals researchers use, and what a product can do with them. The speaker cites the ACL 2025 tutorial on hallucination detection, selective generation and calibration.

The framing question: not "can the LLM say 80%?", but can we rank risky answers, calibrate scores when needed, and decide when silence is better than a plausible answer?

Continue reading — it's free

Google or a one-time email code. No password.