Intel DeepMath Introduces a Smart Architecture to Make LLMs Better at Math

"To address common limitations of LLMs in math reasoning, DeepMath generates small Python scripts that support and enhance its problem-solving process. DeepMath is built on Qwen3-4B Thinking and fine-tuned with GRPO (Group Relative Policy Optimization). Instead of verbose text, the model emits tiny Python snippets for intermediate steps, runs them in a secure sandbox, and folds the results back into its reasoning, reducing errors and output length."

"Based on evaluation across four distinct datasets, MATH500, AIME, HMMT, and HLE, Intel claims that the math agent reduces output length by up to 66% while often improving accuracy, with further performance gains achieved through the use of GRPO. GRPO training introduces rewards for correct answers and for generating code snippets, encourages shorter answers, and varies the temperature during training to promote exploration in the initial training phases and reducing it as the model becomes more proficient."

DeepMath is a lightweight agent built on Qwen3-4B Thinking that solves mathematical problems by emitting small Python executors for intermediate steps. The model runs those snippets in a secure sandbox and incorporates the results back into its reasoning, reducing arithmetic errors and verbose explanations. The agent was fine-tuned with GRPO (Group Relative Policy Optimization), which rewards correct answers and code generation, encourages shorter outputs, and varies temperature during training to balance exploration and proficiency. Evaluation on MATH500, AIME, HMMT, and HLE shows up to 66% shorter outputs and often improved accuracy compared with standard LLM reasoning.

Read at InfoQ