LLMs-the-longer-the-answer-the-better
Although this isn’t fully proven, the longer the answer, the better.
Although we aren’t fully prepared to explain why CoT works, researchers are considering two possible options:
- “Models need tokens to think”: As Andrej Karpathy described in the last Microsoft conference, the more tokens the answer has, the stronger the answer is, as the next word prediction is grounded on more data.
- More computational power: Alternatively, the longer the answer, the more computational effort the model is putting into answering your question. More computational power means better answers.
Bottom line, although this isn’t fully proven, the longer the answer, the better. - ref
Backlinks