Intel's DeepMath slashes AI reasoning bloat by 66% using tiny Python snippets

Intel researchers have built DeepMath, a math agent that trades the long-winded, error-prone text chains of typical LLMs for tiny Python code snippets. The result is a system that can slash output length by up to 66% while often improving accuracy on brutal math benchmarks. It’s a direct shot at the verbose ‘thinking’ traces that make reasoning models slow and expensive.

The project begins with a Qwen3-4B Thinking model and tunes it with Group Relative Policy Optimization, or GRPO. The training explicitly rewards the model for generating short bits of executable code instead of rambling paragraphs. Those snippets are fired off to a locked-down Python sandbox—no file I/O, no network access, strict timeouts—and the computed result is fed back into the model’s context. It’s a clean separation of duties: the LLM handles the high-level reasoning, and a deterministic executor handles the math. As the team puts it, the goal is to “offload deterministic computation to a safe executor.”

On benchmarks like MATH500, AIME, HMMT, and HLE, the fully trained DeepMath agent in agentic mode delivered the highest accuracy with the shortest traces. Ablations tell an interesting story: simply bolting the agent framework onto the untrained Qwen3 model also reduced output length significantly, though with mixed accuracy effects. The magic happens when GRPO training and the agentic inference are combined, creating a model that genuinely prefers concise, computation-driven reasoning.

The training recipe is pragmatic. A linear temperature schedule cools from 1.2 down to 0.7 throughout training to balance early exploration with later precision. The team used 4 in-context examples showing the agent call-and-response pattern, and trained on the Tool-Integrated Reasoning subset of OpenMathReasoning—problems specifically chosen because they benefit from external computation. The code and model are both open, sitting on GitHub and Hugging Face under Intel’s banner. It’s a compelling signal that smaller, specialized models with tight tool integration can punch well above their weight class without drowning you in tokens.