A new study from Towards Data Science reveals hallucinations in large language models (LLMs) stem not from data flaws but from architectural design. Researchers analyzed the residual stream—the internal representation vectors of models like LLaMA-2 13B and Mistral 7B—and found that when models generate confident wrong answers, their internal pathways diverge not from missing data but from deliberate directional shifts in representation space. Gemma 2 2B, despite fewer parameters, showed similar suppression patterns, proving the issue isn't tied to scale.