Paper summary: From Code Foundation Models to Agents and Applications: A Comprehensive Survey and Practical Guide to Code Intelligence

Problem

Code LLMs moved from rule-based systems to Transformers, and from single-digit to over 95% success on benchmarks like HumanEval, powering tools like GitHub Copilot, Cursor and Claude Code. But the field is scattered, and benchmark success doesn’t equal real-world usefulness.

Method

A comprehensive synthesis and practical guide, with the authors’ own analytic and probing experiments, covering the full life cycle:

• data curation and code pre-training
• supervised fine-tuning and reinforcement learning
• prompting paradigms
• autonomous coding agents

It compares general LLMs (GPT-4, Claude, LLaMA) with code-specialised ones (StarCoder, Code LLaMA, DeepSeek-Coder, QwenCoder).

Key results

Beyond the synthesis, the survey articulates a research–practice gap: academic benchmarks versus real development needs, such as correctness, security (see A.S.E from September), awareness of large codebases, and integration with developer workflows.

Limitations

As a survey, it reflects the state of the field at the time of writing.