arXivDaily arXiv每日学术速递 周一至周五更新
arXiv周末暂无论文更新,休息一下吧,周末愉快~~
arXiv 2404.03414cs.CLcs.AI

小型语言模型能否帮助大型语言模型更好地推理?:LM 引导的思维链

Can Small Language Models Help Large Language Models Reason Better?: LM-Guided Chain-of-Thought

Amazon(亚马逊)

Jooyoung Lee, Fan Yang, Thanh Tran, Qian Hu, Emre Barut, Kai-Wei Chang, Chengwei Su

AI总结:

提出 LM-Guided CoT 框架,用经知识蒸馏与强化学习优化的小型 LM 生成推理依据,引导冻结的黑盒大型 LM 推理,在多跳 QA 基准上提升准确率。

摘要:

我们提出了一个新颖框架 LM-Guided CoT,该框架利用轻量级(即 <1B)语言模型(LM)在推理任务中引导黑盒大型(即 >10B)LM。具体而言,轻量级 LM 首先为每个输入实例生成一个推理依据。随后,冻结的大型 LM 被提示基于轻量级 LM 生成的推理依据预测任务输出。我们的方法具有资源效率,因为它只需要训练轻量级 LM。我们通过以下两种方式优化模型:1)知识蒸馏;2)来自面向推理依据和面向任务的奖励信号的强化学习。我们使用多跳抽取式问答(QA)基准 HotpotQA 和 2WikiMultiHopQA 评估该方法。实验结果表明,我们的方法在答案预测准确率方面优于所有基线。我们还发现,强化学习有助于模型生成更高质量的推理依据,并提升 QA 性能。

英文摘要:

We introduce a novel framework, LM-Guided CoT, that leverages a lightweight (i.e., <1B) language model (LM) for guiding a black-box large (i.e., >10B) LM in reasoning tasks. Specifically, the lightweight LM first generates a rationale for each input instance. The Frozen large LM is then prompted to predict a task output based on the rationale generated by the lightweight LM. Our approach is resource-efficient in the sense that it only requires training the lightweight LM. We optimize the model through 1) knowledge distillation and 2) reinforcement learning from rationale-oriented and task-oriented reward signals. We assess our method with multi-hop extractive question answering (QA) benchmarks, HotpotQA, and 2WikiMultiHopQA. Experimental results show that our approach outperforms all baselines regarding answer prediction accuracy. We also find that reinforcement learning helps the model to produce higher-quality rationales with improved QA performance.

机构、总结、中文摘要均由AI分析而来,可能存在错误,请以论文正文为准。

↑