Answer

Ask an LLM the same logic puzzle twice and you might get two different answers. This happens because LLMs generate responses using probabilistic sampling — they don't compute a single correct answer, they sample from a distribution of likely next tokens. Small variations in phrasing or random sampling can push the model down a different reasoning path.

Structural uncertainty is the term researchers use to describe this: the model's reasoning process itself is unstable, not just its surface-level wording. A model might correctly solve a problem one way, then contradict itself when the same problem is reframed.

This is why consistency is increasingly used as a reliability metric. If a model truly understands a logical principle, it should reach the same conclusion regardless of how the question is framed. When it doesn't, that's a sign the model has learned patterns that mimic reasoning rather than the underlying logic itself.