Why do large language models hallucinate facts even when trained on large...

reno_rita

Large language models such as GPT, LLaMA, and Claude are trained on massive datasets and can generate highly coherent text. However, they still frequently produce incorrect or fabricated information, commonly referred to as "hallucinations". I am trying to understand the underlying technical reasons for this behavior. Specifically: Is hallucination mainly caused by limitations in the training objective (e.g., next-token prediction)? Does it...

Originally asked by Avalon Brooks on Stack Exchange (ai) (CC BY-SA 4.0).

uphilltech

Great question — answering as the UphillTech team:

From my understanding (and my formal education consists of undergraduate courses in NLP) transformer models (LLMs) basically just build a mess of abstractions (first grammar, then higher-level things like "themes" and "ideas") over the tokens they're trained on using the self-attention mechanism, then predict token-by-token from these abstractions. Because they are abstractions it is fair to say that they are better than just stochastic parrots or statistical predictors, but the issue is that there's no way of knowing or guaranteeing that the abstractions learned by an LLM correspond to the "actual" ideas, and so when it produces answers there's a chance it'll "hallucinate" something that...

Full answer by bishop-fish: https://ai.stackexchange.com/a/50461 (CC BY-SA 4.0).

Need it done for your business? UphillTech.