Post #1802961
2026-04-08 12:22 UTC
@ngaylinn The "statistically probable" part refers to the pre-trained model, not the fine-tuned one.
One example is asking "when's my birthday". The pre-trained model doesn't know, and so picks a random answer from all possible dates. That's hallucination.
The finetuned model takes that uncertainty, but expresses it as the answer "I don't know".
Sometimes the finetuning fails and we get hallucination in chatbots. This is because they are overtrained to be certain.
Replies (1)
-
@pbloem@sigmoid.social 2026-04-08 12:27
@ngaylinn The innovations here are better finetuning and reasoning traces (that is, having the model output drafts and "reflecting" on them). This is indeed what has reduced the hallucination so much. That's a bit like how you summarize it (hidden prompts), but that doesn't quite capture how rich these reasoning traces can be. See, for example https://arxiv.org/abs/2601.10825 .