A Survey of Vibe Coding with Large Language Models (arxiv.org)

This paper delivers the first comprehensive survey of "Vibe Coding"—a development paradigm where developers validate AI-generated code by observing outcomes instead of reading line-by-line—based on a systematic analysis of over 1,000 papers. The authors formalize Vibe Coding as a Constrained Markov Decision Process that models the triadic interaction among human developers, software projects, and coding agents, and synthesize current practices into five development models: Unconstrained Automation, Iterative Conversational Collaboration, Planning-Driven, Test-Driven, and Context-Enhanced. Crucially, the survey highlights surprising empirical findings: despite high expectations, Vibe Coding can cause productivity regressions and exposes fundamental human–AI collaboration challenges that agent capability improvements alone won’t fix. Technically, the work maps the full Vibe Coding ecosystem—LLMs tailored for coding, LLM-based autonomous agents, agent-aware development environments, and feedback/assessment mechanisms—and argues that success depends on systematic context engineering, robust dev environments, and structured human-agent workflows (e.g., test-driven and planning-driven loops). For practitioners, this means investing in context curation, CI/test pipelines, and feedback channels; for researchers, the formal C-MDP and taxonomy identify open problems in reward/constraint design, agent orchestration, and human-in-the-loop coordination that will shape the next wave of tooling and empirical studies.