Intel DeepMath: Lightweight math reasoning with Qwen3-4B Thinking and sandboxed Python executor

AI Impact Summary

DeepMath pairs Qwen3-4B Thinking with a sandboxed Python executor to emit and evaluate tiny code snippets as part of its reasoning, which reduces verbosity and arithmetic mistakes in math problems. Utilizing smolagents for the agent interface and vLLM as the inference engine, with GRPO fine-tuning that prioritizes correct answers and shorter traces, yields up to 66% shorter outputs and improved accuracy on several math benchmarks. This approach offloads deterministic computation to a safe executor, improving interpretability and potentially lowering per-query cost for math-heavy workloads. Operators should ensure strict sandboxing, per-snippet timeouts, and proper integration testing across target task types beyond the four datasets cited (MATH500, AIME, HMMT, HLE).