A language model answers a question. We change the wording, reverse the order of the options, or add a skeptical reply. Its answer changes. What have we learned?
The tempting conclusion is that the model changed its judgment. But the observation leaves several possibilities open: the judgment changed; the way it was expressed changed; or the scoring rule read the expression differently.
This distinction connects several of my studies of model behavior.
The response format is an experimental variable
In my yes–no bias study, controlled changes to wording and option order separate response preferences from the judgment being measured. A binary answer compresses several things into one token: a word, a position in a list, and a verdict about a proposition.
The multilingual study with Yuwei Lu and Chengkun Cai extends this question across six languages. An opening “yes” or “no” can be read incorrectly when the question is negative. Comparing the sentence-level position with the token reveals discrepancies that would otherwise look like differences in judgment.
The practical lesson is to check what a response affirms before assigning it a score. Agreement between different ways of asking and reading a question is evidence about the measurement itself.
An answer flip has a direction
Social pressure adds another layer. A model may change a wrong answer to a correct one, abandon a correct answer, or hold its position. A flip rate puts the first two events in the same bucket.
In Anatomy of LLM Sycophancy, I compare pressure conditions with matched controls. The wording of the pressure matters: challenging an answer without supplying an alternative need not measure the same behavior as asserting the opposite answer. A useful evaluation must retain that distinction.
What I want from an evaluation
I want to know which interpretation survives controlled comparisons:
- Does a change persist when the question is expressed in an equivalent form?
- Does it appear in the stated position, or only in a token-level score?
- Under pressure, is the change a correction or a capitulation?
- Does the result depend on which items, controls, and pressure conditions were chosen?
These questions make an evaluation more informative. They also make its conclusions more specific. A model can behave consistently along one dimension and remain vulnerable along another; a single aggregate score may conceal both.