Beyond Answers: Transferring Reasoning Capabilities to Smaller LLMs Using Multi-Teacher Knowledge Distillation

TL;DR

TinyLLM, a new knowledge distillation paradigm to learn a small student LLM from multiple large teacher LLMs, and guides the student model to assimilate knowledge from various teacher LLMs, to facilitate the learning of compact language models.

Abstract

Transferring the reasoning capability from stronger large language models (LLMs) to smaller ones has been quite appealing, as smaller LLMs are more flexible to deploy with less expense. Among the existing solutions, knowledge distillation stands out due to its outstanding efficiency and generalization. However, existing methods suffer from several drawbacks, including limited knowledge diversity and the lack of rich contextual information. To solve the problems and facilitate the learning of compact language models, we propose TinyLLM, a new knowledge distillation paradigm to learn a small student LLM from multiple large teacher LLMs. In particular, we encourage the student LLM to not only generate the correct answers but also understand the rationales behind these answers. Given that different LLMs possess diverse reasoning skills, we guide the student model to assimilate knowledge from various teacher LLMs. We further introduce an in-context example generator and a teacher-forcing Chain-of-Thought strategy to ensure that the rationales are accurate and grounded in contextually appropriate scenarios. Extensive experiments on six datasets across two reasoning tasks demonstrate the superiority of our method. Results show that TinyLLM can outperform large teacher LLMs significantly, despite a considerably smaller model size. The source code is available at: https://github.com/YikunHan42/TinyLLM.

Figures (4)

• Figure 1: Pipeline of TinyLLM: Given an input question, we first generate in-context examples and obtain rationales from multiple large LLMs via a teacher-forcing Chain-of-Thought strategy. Later, a small student LLM is trained to integrate rationales from different teachers via multi-task instruction tuning, along with the ground truth label.
• Figure 2: A comparative analysis of TinyLLM against the state-of-the-art Distill-step-by-step method using 80M and 250M FLAN-T5 model architectures across various training set sizes. Dotted line indicates the full fine-tuning (FF) using 100% dataset. It is evident that TinyLLM consistently surpasses the performance of both Distill-step-by-step and full fine-tuning. Notably, TinyLLM achieves this superior accuracy while employing substantially fewer training examples.
• Figure 3: Performance w.r.t. different values of weight $\alpha$.
• Figure 4: Case study of different models' prediction. Examples are selected from the ARC and PIQA datasets. In both cases, TinyLLM successfully generates the correct answer.