CodePlan: Unlocking Reasoning Potential in Large Language Models via Code‑Form Planning

The paper, authored by Tsinghua University graduate Wen Jiaxin and Ant Technology Research Institute associate researcher Guan Jian, investigates the shortcomings of natural‑language‑only reasoning in large language models, such as logical breaks, focus drift, and redundancy.

To overcome these issues, the authors propose CodePlan, a framework that incorporates code‑form planning as an intermediate representation, allowing models to first devise a reasoning blueprint in Python‑style pseudo‑code before expressing the solution in natural language.

CodePlan leverages the rigor of programming constructs—conditional branches, loops, modular functions, and hierarchical architecture—to build precise, reusable reasoning blueprints, eliminating the need for extensive manual annotation by automatically extracting planning signals from existing code data.

Experimental evaluation on 13 challenging benchmarks across five core reasoning tasks shows that CodePlan achieves an average relative performance improvement of 25.1%, with especially large gains on complex multi‑step problems (e.g., over 20 percentage points on the Last Letter task for Mistral‑7B).

The authors also release a dataset of 2 million <prompt, code‑plan, response> triples to facilitate further research, and highlight CodePlan’s benefits for more efficient and stable post‑training of large models.

In conclusion, CodePlan offers a new direction for structured reasoning in AI, bridging the gap between unstructured natural language and the systematic planning required for advanced problem solving, with promising applications in high‑stakes domains such as finance and healthcare.