LLM Honesty Evaluations Affected by Instrument Design.

Evaluating the "honesty" of language models often involves interpreting their outputs as direct evidence of their internal state. However, new research suggests that the design of the evaluation instrument itself can profoundly influence these measurements, rather than solely reflecting the model's inherent honesty. Researchers created a text-adventure game where the game engine, not the language model, held the ground truth about quest completion. A language model played under a budget, eventually declaring its quest status. Crucially, instrument choices, such as expanding the verdict grammar from two to three options or simply disclosing the success criterion, dramatically shifted the model's reported behavior. For instance, disclosing success criteria reduced false verdicts from 18/59 to 0/58. The study highlights that single runs can produce unstable verdict distributions, indicating that reported samples may not represent stable dispositions. It concludes by proposing a four-check integrity protocol for evaluation instruments, emphasizing the need for rigorous design to ensure reliable and auditable assessments of language model honesty.