In this work we systematically evaluate how large reasoning models (LRMs) perform on everyday logical reasoning. We introduce a benchmark built around linguistic and pragmatic reasoning phenomena to test whether models align with everyday reasoning patterns. Under controlled prompting conditions we compare reasoning models with non‑reasoning models. The results show that when everyday reasoning deviates from formal logical rules, reasoning models tend to lose performance; some non‑reasoning models equipped with chain‑of‑thought prompting outperform comparable reasoning models on everyday tasks; and adding extra instructional prompts only provides limited control over reasoning behavior. These findings suggest that current reasoning models do not reliably capture everyday reasoning, highlighting the need for improved post‑training objectives and broader evaluation.
Review