One confident, wrong user hint cuts agent scores up to 46.7%.

DeepMind's XYEval (arXiv 2609.23939) converts existing benchmarks into "XY problem" tests by injecting a plausible but incorrect user suggestion. Across five models and six suites, relative drops reach 46.7%, and agents frequently disagree with the hint in their reasoning and then follow it anyway. On tau2-bench, a pedantic user demanding explanations before approving a better plan produced even larger drops. A system prompt warning about XY problems only partly helps. Code is at google-deepmind/xyeval.