NuminaMath 7B TIR wins 1st AIMO Progress Prize — TIR decoding with Python REPL

AI Impact Summary

NuminaMath 7B TIR achieved first place in the 1st AIMO Progress Prize by leveraging a novel approach combining DeepSeekMath-Base 7B with a tool-integrated reasoning (TIR) decoding algorithm and Python REPL execution. This strategy, built upon a full fine-tuning recipe incorporating Chain of Thought prompting and a synthetic dataset designed for code execution feedback, allowed the model to solve 29 out of 50 problems, demonstrating a significant advancement in AI’s ability to tackle complex mathematical challenges. The use of DeepSpeed and TRL highlights the engineering sophistication behind this winning solution.