A study based on NeurIPS papers found that LLMs were better than human reviewers at predicting which research ideas would be empirically successful