Researchers have found that the specific LLM powering an information-gathering agent changes what it chooses to ask about, even when every other part of the setup is identical.
The study, posted to arXiv, looks at "information elicitation" - the open-ended task of deciding what to ask next as new facts come in, the kind of thing an AI interviewer or research agent has to do constantly. The researchers tested 11 LLMs across different families and sizes, giving each the same information space, the same objectives, and the same rule for picking what to pursue next. That controlled setup let them isolate one variable: how a model judges what information is valuable. They found models diverge in their breadth-versus-depth behavior, and that prior interaction history reshapes how a model evaluates what to ask next. The code, data, and interaction logs are posted on GitHub.
This matters because a lot of agent products quietly assume the model is a neutral executor of a strategy, when the strategy is actually baked into the model itself. Swap the underlying LLM in a research agent, customer-interview bot, or diagnostic tool, and you may get a systematically different interviewer - broader but shallower, or narrower but more thorough - without changing a single line of prompt.
It is a useful reminder that "just swap the model" is rarely free. The paper does not name winners or losers, but it gives builders a way to test which model actually asks the right questions for their use case, rather than assuming any sufficiently capable LLM will do.