This survey sets out a roadmap for Modern Code Review (MCR), which it describes as a cornerstone of software quality assurance and a channel for knowledge transfer that nonetheless remains cognitively demanding and resource-intensive. It collects studies from Web of Science, the ACM Digital Library and IEEE Xplore whose titles or topics mention “code review” or “code inspection”, starting from 2013, which the authors identify as the year of the first MCR study, and ending in November 2025. From 770 initial hits, deduplication, inclusion criteria and snowballing leave 327 primary studies.
The authors split the literature into improvement techniques (153 papers, 46.8%), which automate or optimize specific review tasks — code change analysis, reviewer recommendation, review comment synthesis, review comment analysis and unified automation frameworks — and understanding studies (174 papers, 53.2%), which examine quality assurance and reliability, process efficiency and workflow patterns, human factors and social interaction, and the evolution toward human-AI collaboration. They argue that large language models are shifting MCR from passive tool support toward a potential human-AI symbiotic partnership, but also bring new risks, and use a SWOT analysis to diagnose the gap between AI capabilities and industrial realities.
Key Points
- Publication has grown in three stages: fewer than 10 papers a year in 2013–2014, 14–22 a year in 2015–2021, and over 40 a year in 2022–2025; 209 of the 327 studies (about 64%) appeared since 2021, a period earlier surveys did not cover.
- Conference papers make up 65% of the corpus (213 papers), spread over 68 venues, with ICSE the most frequent single venue and Empirical Software Engineering the leading journal.
- The SWOT analysis names three strengths — unified multi-task frameworks, LLMs’ generative proficiency, and emerging multi-agent systems — and three weaknesses: a context gap, since models work on local snippets or file-level diffs without project history, issue trackers or design documents; hallucination and trust problems, including high false-positive rates and superficial “LGTM smell” approvals; and misaligned metrics such as BLEU that do not capture the logic, constructiveness or educational value of feedback.
- As opportunities it names industrial data flywheels, specialized domains such as security auditing, education and green software engineering, and process integration into CI/CD pipelines and IDEs; as threats, erosion of collective ownership, deskilling of junior developers, and amplification of biases such as gender bias or toxic tone present in training data.
- The authors propose three paradigm shifts: from passive assistants to proactive collaborators, with repository-specific long-term memory via retrieval-augmented generation, multi-agent collaboration and continuously updated knowledge; from accuracy metrics to value-driven evaluation of utility, acceptability, cognitive load and verification cost, on benchmarks reflecting multi-file, multi-turn review; and from automation to human-centric symbiosis, with human sign-off for critical changes, AI acting as a mentor, and bias detection.
- They envision a move from asynchronous, web-based pull-request gatekeeping to IDE-native agents that give context-aware guidance while code is being written, which they call a “pre-emptive review” model.
Notes
In the survey’s account of human-AI collaboration, cited studies report that LLM-assisted reviews can increase verification overhead and pull-request closure time, bias reviewers toward lower-severity issues, receive less visual attention when code is AI-generated, and fail to foster the collective accountability that human review creates — findings the authors use to argue for hybrid workflows that keep human oversight. They describe understanding studies as the “compass” and improvement techniques as the “engine” of the field, each driving new work in the other. The taxonomy and its paper lists are published in a GitHub repository. For a single early-stage multi-agent review system, see AI-powered Code Review with LLMs: Early Results; related concepts in this wiki include Agentic Code Review and Context-aware code review.