“There's an interesting new paper about this problem: Think before you speak”
Think before you speak: Training Language Models With Pause Tokens
在输入后追加可学习的暂停词元,让模型回答前多算几步
语言模型通常逐个生成词元,作者让模型在输出第 K+1 个词元前处理更多隐藏向量:在输入末尾追加一串可学习的 pause 词元,并延迟提取输出到最后一个 pause 词元之后。在 1B 和 130M 参数、于 C4 上预训练的 decoder-only 模型上,只有预训练和微调都使用 pause 时推理延迟才有收益;1B 模型在 9 个任务中 8 个有提升,其中 SQuAD 的 EM 提升 18%,CommonSenseQA 提升 8%,GSM8k 准确率提升 1%。
HN 上的人把它当作「给模型更多思考计算」的例子,引用它说明即使是不含新信息的占位词元也能提升表现,并与思维链等做对照。
HN 书评
推荐在前,同类按时间 · 每篇最多保留 20 条,统计用全部提及“Pause tokens (thinking tokens) are also an interesting method to achieve that and seems to have a positive effect on performance:”
“One idea I was interested in was after reading the paper on introducing pause tokens[1]”
“You can even even implement a computation token if you wish.”
“even if it's just dummy tokens without any new information, it works.”
“There's a recent paper which seeks to explicitly perform time-to-think using pause tokens[1].”
看全部 14 条 →
“let it be pause tokens”
“You could draw parallels to allowing LLMs to emit pause tokens to get more time to think (Think before you speak and similar).”
“are directly dependent on the model recognizing whether the answer it arrived at is one it should”
“So yes you can increase compute length arbitrarily by increasing tokens.”
“There's been publications with a pause token, backspace token ([OTHER_PAPER]), or a think token ([OTHER_PAPER])”
“Of course, training the model to do this (use pause tokens) on purpose works too: Think before you speak”
“"Think before you speak: Training Language Models With Pause Tokens" - Think before you speak”
“You can train a special token for this purpose [2].”
- Think Before You Speak: Training Language Models with Pause Tokens ↗og_kalu · 2023-10-04 · 2 分