# STaR: Bootstrapping Reasoning With Reasoning

**Authors:** Eric Zelikman, Yuhuai Wu, Jesse Mu, Noah D. Goodman  
**arXiv:** [2203.14465](https://arxiv.org/abs/2203.14465)  
**Date:** 2022-03-28  
**Summary page:** https://www.summarizepaper.com/en/paper/2203.14465v2/

## Key points

- Authors discuss the importance of generating step-by-step "chain-of-thought" rationales to enhance language model performance on intricate reasoning tasks
- Challenges in inducing language model rationale generation include constructing extensive datasets or compromising accuracy with few-shot inference methods
- Proposal of a novel technique called "Self-Taught Reasoner" (STaR) to bootstrap the model's ability for complex reasoning tasks
- STaR involves iteratively leveraging a small number of rationale examples along with a large dataset lacking rationales
- Approach includes generating rationales, fine-tuning based on successful rationales, and repeating the process to enhance reasoning abilities
- STaR significantly enhances performance across various datasets compared to models directly fine-tuned for predicting final answers
- Achieves comparable results to fine-tuning a state-of-the-art language model that is 30 times larger on CommensenseQA tasks

## Summary

, , , , 

In their paper titled "STaR: Bootstrapping Reasoning With Reasoning," authors Eric Zelikman, Yuhuai Wu, Jesse Mu, and Noah D. Goodman discuss the importance of generating step-by-step "chain-of-thought" rationales to enhance language model performance on intricate reasoning tasks such as mathematics and commonsense question-answering. They highlight the current challenges in inducing language model rationale generation, which typically involve either constructing extensive rationale datasets or compromising accuracy by relying solely on few-shot inference methods. To address these challenges, the authors propose a novel technique called the "Self-Taught Reasoner" (STaR), which aims to iteratively leverage a small number of rationale examples along with a large dataset lacking rationales. The key idea behind STaR is to bootstrap the model's ability to engage in progressively more complex reasoning tasks. The approach involves a simple loop: generate rationales to answer multiple questions based on a few initial rationale examples; if the generated answers are incorrect, attempt to generate a rationale given the correct answer; fine-tune the model on all successful rationales that led to accurate answers; and repeat this process. The authors demonstrate that STaR significantly enhances performance across various datasets compared to models directly fine-tuned for predicting final answers. Moreover, STaR achieves comparable results to fine-tuning a state-of-the-art language model that is 30 times larger on CommensenseQA tasks. This highlights how STaR enables a model to enhance its own reasoning abilities by learning from the reasoning it generates during training iterations. Overall, the research presented in this paper sheds light on an innovative approach for improving language models' reasoning capabilities through self-learning mechanisms, showcasing promising results in enhancing performance on complex reasoning tasks without requiring extensive manual annotation efforts or sacrificing accuracy.