批评xyz7786 · 2025-03
“they basically just say "answer in X words" and leave it at that”
批评论文基本只是要求“用 X 个词回答”,别无更多
简洁思维链(CCoT)提示:回答缩短约 49%,选择题准确率基本不变
论文提出简洁思维链(CCoT)提示,并在 GPT-3.5 和 GPT-4 上用多项选择问答基准比较标准 CoT 与 CCoT。CCoT 使两个模型的平均回答长度减少 48.70%,对解题表现影响可忽略,平均每 token 成本下降 22.67%;但在数学题上,GPT-3.5 使用 CCoT 会损失 27.69% 的表现。
评论者用它支持「简洁提示能大幅缩短回答而不总是损害质量,但依任务而异」,还提到某个「Caveman」提示借鉴了它;也有人批评它基本只是限制回答字数。
“they basically just say "answer in X words" and leave it at that”
“There is actual research suggesting concise prompting can reduce response length substantially without always wrecking quality, though it is task-dependent and can hurt in some domains.”
“It builds on the same concise-CoT research behind the original Caveman prompt”