软件工程：2025–2026 年度综述

An annual survey of the cs.SE arXiv category, covering research published from September 2025 – August 2026. Built on a close-reading set of 100 papers and organised into 40 chapters, it maps the year by the field's research sub-directions.

Book framework (share of the year's output):

• Data-Driven Systems and Machine Learning for Safety (19.5%) — Applies machine learning and data-driven models to build and analyse software systems, with particular attention to safety-critical behaviour.
• Large Language Models for Software Engineering and Specification (14.6%) — Applies large language models to program analysis, specification, and requirements engineering to improve code quality.
• Agentic LLMs: Tools, Skills, and Multi-Agent Execution (10.8%) — Investigates LLM agents that call external tools, acquire skills, and coordinate multi-agent execution to carry out complex software tasks.
• Open Source Software Security and Vulnerability Detection (9.7%) — Examines security practices, dependency management, and automated detection of vulnerabilities in open-source software.
• Agentic AI and Generative AI Assisted Human Software Development (8.4%) — Examines how agentic and generative AI assists human developers across requirements, design, coding, and collaborative workflows.
• Automated Test Generation and Metamorphic Testing (7.2%) — Develops methods to automatically generate test cases and apply metamorphic testing, especially for validating LLM-based systems.
• Software Engineering Research and Development (6.4%) — Covers empirical software engineering research, developer workflows, and the practices and tools that shape modern development.
• Agentic Coding and Software Engineering Benchmarks (5.9%) — Evaluates agentic coding assistants through benchmarks that measure their ability to solve realistic software engineering tasks.
• Large Language Models for Code Generation and Automated Review (5.4%) — Studies how large language models generate source code and automate code review, evaluating the correctness, quality, and maintainability of the code they produce.
• Automated Program Repair and Bug Localization (5.1%) — Studies automated approaches to localise faults and generate patches that repair faulty programs.
• Repository-Level Retrieval Augmented Generation (5.0%) — Uses repository-level retrieval-augmented generation to supply coding agents with relevant context from large codebases.
• Quantum Software Engineering and Classical Hybrid Computing (1.8%) — Explores how to write, test, and verify quantum programs and integrate them with classical computing in hybrid systems.

Each sub-direction is covered across three chapters — how it framed its questions, the strength of its key findings, and its open debates and frontiers — so the reader gets a traceable picture of what this part of the field actually produced over the year.