Large language models often produce plausible yet inaccurate outputs and self-assess their confidence by repeated querying, which can miss overconfidence errors. To address this, MIT researchers developed an uncertainty quantification method comparing a model’s output to that of peer models, measuring disagreement rather than self-consistency alone.
Combining cross-model disagreement with self-consistency yielded a total uncertainty metric that outperformed existing methods across ten tasks including question-answering and math reasoning, providing a more reliable indicator of when LLM responses may be erroneous, critical for high-stakes applications such as healthcare or finance.
Source & further reading
A better method for identifying overconfident large language models ↗Source published 2026-03-19. This report was prepared with AI assistance from the linked primary-source material. Claims and limitations should be checked against the original announcement or research.