Are Large Reasoning Models Good Everyday Reasoners? A Comparison with Non-Reasoning Models on Logical Reasoning Tasks

Explain Like I'm Normal

Researchers tested 'reasoning models' (like o1) against standard LLMs on everyday logical tasks that require social context and pragmatics rather than formal math. The study found that extra test-time compute actually hurts performance when the logic relies on human nuance rather than textbook rules. Essentially, these models are becoming 'super-geniuses' who lack basic common sense.