Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Code to Think, Think to Code: A Survey on Code-Enhanced Reasoning and Reasoning-Driven Code Intelligence in LLMs

Dayu Yang, Tianyang Liu, Daoan Zhang, Antoine Simoulin, Xiaoyi Liu, Yuwei Cao, Zhaopu Teng, Xin Qian, Grey Yang, Jiebo Luo, Julian McAuley

TL;DR

<3-5 sentence high-level summary> The paper surveys the bidirectional relationship between code and reasoning in LLMs, arguing that executable and structured code provides a robust medium for reasoning while enhanced reasoning expands the scope and quality of code intelligence. It presents a taxonomy and synthesis of approaches spanning executable program-based reasoning (PoT/PaL), dynamic code-language integration, non-executable code representations, and code-driven training, as well as reasoning-driven code generation, code understanding, interactive programming, and autonomous code agents. It also reviews benchmarks and evaluation frameworks (e.g., CRUXEval, CodeMMLU, SWE-bench) and identifies core challenges—interpretability, scalability, data quality, and safe tool usage—offering future directions like code-language hybrids, modular agents, and RL-based improvements. The work aims to guide researchers and practitioners toward more interpretable, reliable, and capable code-centric AI systems with stronger reasoning faculties.

Abstract

In large language models (LLMs), code and reasoning reinforce each other: code offers an abstract, modular, and logic-driven structure that supports reasoning, while reasoning translates high-level goals into smaller, executable steps that drive more advanced code intelligence. In this study, we examine how code serves as a structured medium for enhancing reasoning: it provides verifiable execution paths, enforces logical decomposition, and enables runtime validation. We also explore how improvements in reasoning have transformed code intelligence from basic completion to advanced capabilities, enabling models to address complex software engineering tasks through planning and debugging. Finally, we identify key challenges and propose future research directions to strengthen this synergy, ultimately improving LLM's performance in both areas.

Code to Think, Think to Code: A Survey on Code-Enhanced Reasoning and Reasoning-Driven Code Intelligence in LLMs

TL;DR

<3-5 sentence high-level summary> The paper surveys the bidirectional relationship between code and reasoning in LLMs, arguing that executable and structured code provides a robust medium for reasoning while enhanced reasoning expands the scope and quality of code intelligence. It presents a taxonomy and synthesis of approaches spanning executable program-based reasoning (PoT/PaL), dynamic code-language integration, non-executable code representations, and code-driven training, as well as reasoning-driven code generation, code understanding, interactive programming, and autonomous code agents. It also reviews benchmarks and evaluation frameworks (e.g., CRUXEval, CodeMMLU, SWE-bench) and identifies core challenges—interpretability, scalability, data quality, and safe tool usage—offering future directions like code-language hybrids, modular agents, and RL-based improvements. The work aims to guide researchers and practitioners toward more interpretable, reliable, and capable code-centric AI systems with stronger reasoning faculties.

Abstract

In large language models (LLMs), code and reasoning reinforce each other: code offers an abstract, modular, and logic-driven structure that supports reasoning, while reasoning translates high-level goals into smaller, executable steps that drive more advanced code intelligence. In this study, we examine how code serves as a structured medium for enhancing reasoning: it provides verifiable execution paths, enforces logical decomposition, and enables runtime validation. We also explore how improvements in reasoning have transformed code intelligence from basic completion to advanced capabilities, enabling models to address complex software engineering tasks through planning and debugging. Finally, we identify key challenges and propose future research directions to strengthen this synergy, ultimately improving LLM's performance in both areas.

Paper Structure

This paper contains 94 sections, 3 figures, 5 tables.

Table of Contents

  1. Introduction
  2. Code-enhanced Reasoning
  3. Generating as Code Aids Reasoning
  4. Single Execution
  5. Dynamic Code-Language Integration
  6. Non-Executable Program Representations
  7. Training with Code
  8. Empowering Reasoning and Planning Through Code Training
  9. Training Strategies Based on Code
  10. Code-only Training Strategies
  11. Hybrid-data Training Strategies
  12. Reasoning-Enhanced Code Intelligence
  13. Essential Code Intelligence
  14. Integration of Reasoning Capabilities
  15. Reasoning for Code Generation
  16. ...and 79 more sections

Figures (3)

  • Figure 1: Bidirectional enhancement between code properties and reasoning capabilities.
  • Figure 2: Taxonomy of interplay between Code and Reasoning.
  • Figure 3: Full taxonomy illustrating the interplay between code and reasoning.