Hallucinations in LLMs Are Not a Bug in the Data

Hallucinations in LLMs are presented as a structural feature of their architecture rather than a data or training bug. By analyzing the internal trajectory of a model's residual stream, it's shown that during a hallucination, the model actively suppresses the correct answer instead of simply failing to retrieve it. A metric called the "commitment ratio" reveals that probability mass is moved away from the correct token, indicating a conflict where the model prioritizes contextual coherence over factual accuracy. This behavior is an emergent property of optimizing for next-token prediction, and while this geometric signature can be used for detection, it suggests that such monitoring tools must be domain-specific.