# T-SciQ: Teaching Multimodal Chain-of-Thought Reasoning via Large Language Model Signals for Science Question Answering

**Authors:** Lei Wang, Yi Hu, Jiabang He, Xing Xu, Ning Liu, Hui Liu, Heng Tao Shen  
**arXiv:** [2305.03453](https://arxiv.org/abs/2305.03453)  
**Date:** 2023-05-05  
**Summary page:** https://www.summarizepaper.com/en/paper/2305.03453v2/

## Key points

- Large Language Models (LLMs) have shown remarkable performance in various NLP tasks, including complex chain-of-thought (CoT) reasoning.
- Collecting high-quality CoT rationales for training these models is laborious and costly.
- The proposed T-SciQ method aims to teach science question answering with LLM signals.
- T-SciQ generates high-quality CoT rationales as teaching signals and trains smaller models for CoT reasoning in complex multimodal scenarios.
- T-SciQ introduces a novel data mixing strategy to produce more effective teaching data samples.
- Experimental results on the ScienceQA benchmark show that T-SciQ achieves a new state-of-the-art performance with an accuracy of 96.18%.
- T-SciQ outperforms the most powerful fine-tuned baseline by 4.5%.
- This work focuses on mixing single and multi-step reasoning chains based on validation set predictions, setting it apart from previous studies.

## Summary

Recently, Large Language Models (LLMs) have shown remarkable performance in various Natural Language Processing (NLP) tasks, including complex chain-of-thought (CoT) reasoning. However, collecting high-quality CoT rationales for training these models is a laborious and costly process. To address this issue, the proposed T-SciQ method aims to teach science question answering with LLM signals. It generates high-quality CoT rationales as teaching signals and trains smaller models to perform CoT reasoning in complex multimodal scenarios. The approach also introduces a novel data mixing strategy to produce more effective teaching data samples. Experimental results on the ScienceQA benchmark demonstrate that T-SciQ achieves a new state-of-the-art performance with an accuracy of 96.18%, outperforming the most powerful fine-tuned baseline by 4.5%. Additionally, this work focuses on mixing single and multi step reasoning chains based on validation set predictions, setting it apart from previous studies.