A team at the University of Florence published arXiv paper 2606.07515, putting 8 SOTA large models (16 configurations with/without CoT) through discrete probability reasoning problems that are "formally provable and can trigger intuitive bias." The result is a watershed: the standard question averages 0.96 (9 of 16 exceed 0.99), and the counter-intuitive variants drop directly to 0.59, with the strongest ChatGPT 5.4 Thinking only reaching 0.84. The paper then conducts three "dimensionality reduction attacks": rewriting the question to an isomorphic but unfamiliar version drops accuracy by 20%; planting "reasonable-looking" wrong answers generated by other models in the prompt drops performance by up to 34%, with no model immune; the most unusual is Mistral Large 3, which gains almost nothing from turning on CoT. The conclusion is direct: today's LLMs are not probabilistic reasoners, but "probability question repeaters" in the training corpus. Their robustness on standard questions comes more from retrieving the question template than from internal verification of probability axioms; the "pleasing" training of the RLHF stage also locks the reasoning ceiling on "question feel." This exactly explains why recent works like GRPO and On-Policy Distillation have started pushing the correction pressure from the result layer to the rollout layer — the pure-question-feel dividend is about to end.