Self-consistency probing

How it works

A model that gives different answers to the same question on different runs is unreliable regardless of which particular answer happens to be right — inconsistency alone is the finding. Self-consistency probing samples the same prompt many times (varying only the seed or temperature) or asks logically equivalent rephrasings of the same question, then measures how often the answers agree. High agreement across samples turns raw stochasticity into a quantitative reliability signal for a given prompt or task; disagreement pinpoints exactly the questions where the model doesn't reliably 'know' the answer, as distinct from questions it simply gets wrong every time. It's a cheap, oracle-free way to separate 'the model doesn't know' from 'the model consistently believes something incorrect'.

When to use it

Reliability assessment on questions where an answer key is expensive or unavailable; flagging low-confidence outputs for human review or escalation; calibration studies comparing stated confidence against actual agreement rate.

Limitations

Consistency is not correctness — a model can converge on the same wrong answer every time, so a high agreement rate alone tells you nothing about accuracy and needs to be paired with an oracle method on at least a sample. Needs enough repeated samples to be statistically meaningful, which adds cost proportional to how many are drawn.

Cite this