GAIR/LIMO
LIMO (Less Is More for Reasoning) is a 32.8 billion parameter instruction-tuned causal language model developed by GAIR, built upon the Qwen2.5-32B-Instruct backbone. It specializes in mathematical reasoning, achieving state-of-the-art performance on complex math benchmarks like AIME24 and MATH500 with significantly less, high-quality training data. This model demonstrates strong generalization across diverse mathematical problem types, making it suitable for advanced reasoning tasks.
Loading preview...
LIMO: Less Is More for Reasoning
LIMO is a 32.8 billion parameter model developed by GAIR, fine-tuned on the Qwen2.5-32B-Instruct backbone, that redefines mathematical reasoning performance. It achieves state-of-the-art results using a highly curated, small dataset of only 817 training samples, challenging the traditional reliance on massive datasets for superior performance.
Key Capabilities & Performance
- Exceptional Mathematical Reasoning: LIMO demonstrates significant improvements across 10 mathematical benchmarks.
- Achieves 57.1% on AIME24, a +50.6% improvement over previous SOTA (6.5%).
- Scores 94.8% on MATH500, a +35.6% improvement over previous SOTA (59.2%).
- Outperforms previous state-of-the-art models by substantial margins on AMC23, OlympiadBench, CHMath, Gaokao, Kaoyan, and GradeSchool benchmarks.
- Efficient Training: Achieves high performance with a remarkably small, high-quality training dataset, highlighting the effectiveness of data curation over sheer volume.
- Strong Generalization: Shows robust performance across a diverse range of mathematical problem types.
When to Use LIMO
- Advanced Mathematical Problem Solving: Ideal for applications requiring precise and complex mathematical reasoning.
- Educational Tools: Can be integrated into platforms for generating solutions or explanations for challenging math problems.
- Research in Efficient LLM Training: Useful for researchers exploring data-centric approaches to improve model performance with fewer resources.
LIMO's approach demonstrates that strategic data curation can lead to highly effective models, particularly in specialized domains like mathematical reasoning.