DeepSeek-R1 Math Proofs With Symbolic Verification

DeepSeek-R1-Zero and R1 were trained using GRPO and symbolic rewards based on the accuracy and format of generations. Pair math-style sampling (temperature=0.7, top_p=1.0) with a clear final format delimiter.

API Config: temperature=0.7, top_p=1.0 User Prompt: Prove that for every positive integer n, the quantity n^5 - 5n^3 + 4n is divisible by 120. State the final verified divisor inside <verified_answer> tags.