LIMO: Less is More for Reasoning

arxiv:2502.03387

Original source · Paper JSON

Reasoning and the "physics" of language models · 2025-02-05

Cites (4) · Cited by (4)

Complete cleaned Markdown · Source Markdown including bibliography

Read in parts

1. Part 1
2. Part 2

Contents

• LIMO: Less is More for Reasoning
• Abstract
• 1 Introduction
• 2 Related Work
• 2.1 Evolution of Mathematical Reasoning in LLMs
• 2.2 Test-time Scaling and Long Chain Reasoning
• 2.3 Data Efficiency in Language Models
• 3 LIMO Dataset
• 3.1 High-Quality Data Curation
• 3.1.1 Question Selection
• 3.1.2 Reasoning Chain Construction
• 4 Training Recipe
• 5 Evaluation Framework
• In-domain Evaluation
• Out-of-distribution Evaluation
• Performance metrics
• 6 Experiment
• 6.1 Baselines
• 6.2 Main Results
• In-domain Performance
• Out-of-domain Generalization
• Comparison with Larger Datasets
• Overall Performance
• 6.3 Analysis
• 6.3.1 RQ1: Impact of Reasoning Chain Quality
• 6.3.2 RQ2: Impact of Question Quality
• 6.3.3 RQ3: LLM Backbone - Pre-trained Knowledge
• 6.3.4 RQ4: LLM Backbone - Model Size
• 6.3.5 RQ5: Sample Efficiency
• 7 Conclusion
• 8 Acknowledgement