Large Language Models often rely on semantic approximations rather than robust symbolic execution when performing deterministic tasks. To address this limitation, researchers introduced MIMIC, a framework designed to leverage executable code as a rigorous medium for reasoning data synthesis.
The core mechanism involves transforming algorithms into verifiable reasoning trajectories through narrative fusion and dynamic code instrumentation. This process generates explicit intermediate execution states that form a Code-Instrumented Reward (CIR). Unlike traditional methods requiring external reward models, CIR provides dense, high-fidelity process supervision directly from the code's execution flow.
Extensive evaluations demonstrate that models trained via Supervised Fine-Tuning (SFT) and Group Relative Policy Optimization (GRPO) on the synthesized dataset achieve substantial gains. The improvements are consistent across general reasoning tasks, complex mathematical benchmarks, and fine-grained deterministic problems.
The study concludes that the procedural rigor of executable code effectively unlocks generalized reasoning capabilities in LLMs. Researchers have made the associated code and data available for public use.
Source: https://arxiv.org/abs/2609.16076
