목차

• Minerva: Solving Quantitative Reasoning Problems with Language Models
  • Abstract
  • 1. Introduction
    • 1.1 Our Contribution
    • 1.2 Related Works
  • 2. Training and Evaluation
    • 2.1 Mathematical Training Dataset
    • 2.2 Models and Training Procedure
    • 2.3 Evaluation Datasets
    • 2.4 Undergraduate-Level STEM Problems
    • 2.5 Inference-Time Techniques
  • 3. Results
    • 3.1 Basic arithmetic
  • 4. Performance Analysis
    • 4.1 Model Mistakes
    • 4.2 False Positives
  • 5. Memorization
    • 5.1 Training and Evaluation Dataset Overlap
    • 5.2 Performance on Modified MATH Problems
    • 5.3 BLEU Score Between Ground Truth and Generated Solutions
  • 6. Conclusions and Discussion
    • 6.1 Limitations of Our Approach
    • 6.2 Societal Impact

이 글에서는 Google Inc.에서 발표한 Minerva: Solving Quantitative Reasoning Problems with Language Models 논문을 간략하게 정리한다.

Minerva: Solving Quantitative Reasoning Problems with Language Models

논문 링크: Minerva: Solving Quantitative Reasoning Problems with Language Models

Blog: https://ai.googleblog.com/2022/06/minerva-solving-quantitative-reasoning.html

• 2022년 6월(Arxiv)
• Aitor Lewkowycz, Anders Andreassen, David Dohan et al.

Abstract

1. Introduction

1.1 Our Contribution

1.2 Related Works

2. Training and Evaluation

2.1 Mathematical Training Dataset

2.2 Models and Training Procedure

2.3 Evaluation Datasets

2.4 Undergraduate-Level STEM Problems

2.5 Inference-Time Techniques

3. Results

3.1 Basic arithmetic

4. Performance Analysis

4.1 Model Mistakes

4.2 False Positives

5. Memorization

5.1 Training and Evaluation Dataset Overlap

5.2 Performance on Modified MATH Problems

5.3 BLEU Score Between Ground Truth and Generated Solutions

6. Conclusions and Discussion

6.1 Limitations of Our Approach

6.2 Societal Impact