The Imitation Game: When LLMs Learn to Reason Like Programs via Code-Centric Reasoning Data Synthesis

Researchers proposed MIMIC, a framework that uses executable code as a medium for reasoning data synthesis, transforming algorithms into verifiable reasoning trajectories through narrative fusion, code-guided test synthesis, and dynamic code instrumentation. According to the arXiv paper 2609.16076v1, the explicit intermediate execution states form a Code-Instrumented Reward (CIR) that provides dense process supervision for reinforcement learning without external reward models. Models trained via SFT and GRPO on the synthesized dataset showed substantial, consistent accuracy gains across general reasoning, complex mathematical benchmarks, and fine-grained deterministic tasks, with code and data available at https://github.com/zjy1298/MIMIC.

arXiv:2609.16076v1 Announce Type: new Abstract: Large Language Models LLMs excel at programming tasks but frequently fail at deterministic, fine-grained reasoning in natural language, relying heavily on semantic approximations rather than robust symbolic execution. To bridge this gap, we propose MIMIC, a framework that leverages executable code as a rigorous medium for reasoning data synthesis. MIMIC fundamentally transforms algorithms into verifiable reasoning trajectories through narrative fusion, code-guided test synthesis, and dynamic code instrumentation. Crucially, these explicit intermediate execution states naturally form a Code-Instrumented Reward CIR , providing dense, high-fidelity process supervision for reinforcement learning without external reward models. Extensive evaluations reveal that models trained via SFT and GRPO on our synthesized dataset achieve substantial, consistent gains. Our method significantly elevates accuracy across general reasoning, complex mathematical benchmarks, and fine-grained deterministic tasks, demonstrating that the procedural rigor of executable code can effectively unlock and enhance the generalized reasoning capabilities of LLMs. Our code and data are available at https://github.com/zjy1298/MIMIC.