How would you design a reward function for a math RL tr...

Design the reward function using binary accuracy based on symbolic equivalence (SymPy) rather than simple string matching, supplemented by a lightweight format reward to enforce response structure. Filter problems to ensure a 20%-80% group pass-rate window. Avoid multiple-choice or true/false formats, as these allow the model to achieve high rewards through random-walking rather than actual reasoning.