How LLMs Learn to Solve Complex Math

Large Language Models excel in many tasks but often fail at multi-step reasoning, especially in mathematics. This paper introduces a novel arithmetical puzzle benchmark and a synthetic data pipeline to fine-tune open-llama-3B. Experiments show significant improvements in zero-shot accuracy across in-domain and out-of-domain datasets, suggesting that high-quality synthetic data can help LLMs generalize better in complex reasoning tasks.