This OpenAI paper has some very compelling explanations on why hallucinations persist and what to do about it. Specifically in post-training, we should stop penalizing models for saying that they don’t know an answer, and stop rewarding overly confident guessing. Eradicating https://t.co/lyByay49ul