Wizardmath
WizardMath is a mathematical specialist model created by applying Evol-Instruct (iterative problem hardening) to Llama-2 with process reward modeling that rewards correct reasoning steps, not just correct answers, pioneering the combination of evolving problem difficulty with step-by-step reasoning supervision to achieve state-of-the-art mathematical reasoning.
Dual Innovation: Evol-Instruct + Process Supervision
| Component | Function | Impact |
|---|---|---|
| Evol-Instruct | Automatically increase problem difficulty progressively | From simple arithmetic to competition-level reasoning |
| Process Rewards | Reward correct steps individually, not just final answers | Models learn genuine reasoning, not answer shortcuts |
Evol-Instruct for Math: Takes simple math problems and uses an AI agent to systematically make them harder:
1. "1+1" → "Write Python to add two numbers"
2. → "Multi-threaded C++ matrix addition"
3. → "Optimize for cache coherency"
Models trained on this progression learn deep problem-solving skills.
Process Reward Model: Instead of "Is final answer correct? Yes/No", WizardMath uses "Is step 1 correct? +1. Is step 2 correct? +1."—teaching rigorous reasoning at every stage, not just answer-matching.
Performance: 81.6% on GSM8K—exceeds ChatGPT (80.8%), achieves 22.7% on MATH—significantly above base Llama (13.5%), proving process-based rewards matter more than outcome rewards.
Impact on the Field: The Evol-Instruct + Process Reward combination became a template for domain-specialized improvements: applied to code (WizardCoder), general tasks (WizardLM), and scientific reasoning—establishing it as one of the most broadly influential fine-tuning methodologies in open-source AI.