Answer

Even when asking the same question, frontier LLMs often disagree — sometimes significantly. This happens because each model is trained on different data, uses different training objectives, and applies different fine-tuning and reinforcement signals that shape what the model treats as "correct."

Factual knowledge in LLMs isn't stored like a database lookup. It's compressed into billions of numerical weights during training, meaning facts are distributed across the model in a fuzzy, probabilistic way. Small differences in training data or methods can cause models to reach different conclusions.

This is one reason why model disagreement is actually a useful signal. When multiple models align on an answer, confidence is higher. When they diverge, it's a strong hint that the question is contested, underrepresented in training data, or that at least one model is hallucinating.