INTELLECT-MATH: Frontier Mathematical Reasoning through Better Initializations for Reinforcement Learning
INTELLECT-MATH is a 7B parameter model optimized for mathematical reasoning. It was trained in two stages, an SFT stage, in which the model was fine-tuned on verified QwQ outputs, and an RL stage, in which the model was trained using the PRIME-RL recipe.
We demonstrate that the quality of our SFT data can impact the performance and training speed of the RL stage: Due to its better synthetic SFT dataset that encourages the model to imitate the reasoning behavior of a strong teacher model, INTELLECT-MATH outperforms Eurus-2-PRIME, the previous state-of-the-art trained with PRIME-RL, and matches its performance with 10x faster training.
| Intellect-Math (Step 255) | Intellect-Math (Step 47) | Eurus-2-Prime (Step 592) | Intellect-Math-SFT | Eurus-2-SFT | Qwen-2.5-Math | |
|---|---|---|---|---|---|---|
| MATH-500 | 82.0 | 81.6 | 79.2 | 72.8 | 65.1 | 79.8 |
| OLYMPIADBENCH | 49.5 | 46.7 | 42.1 | 39.1 | 29.8 | 40.7 |
| AIME 2024 | 26.7 | 26.7 | 26.7 | 16.6 | 3.3 | 13.3 |
| AMC | 60.2 | 57.8 | 57.8 | 45.8 | 30.1 | 50.6 |
| MINERVA MATH | 39.7 | 37.8 | 38.6 | 33.8 | 32.7 | 34.6 |
| AVG | 51.6 | 50.1 | 48.9 | 41.6 | 32.2 | 43.8 |