AI code research pays less attention to efficiency and maintainability

What they studied

The authors reviewed research on how code generated by large language models is evaluated and improved. Their central question was whether existing research covers the range of quality standards needed for professional software development, rather than concentrating on only some aspects of quality.

To organize the review, they adapted a software quality framework from the ISO/IEC 25010 standard. It covers four attributes: functional correctness, security, performance efficiency, and maintainability. In plain terms, these concern whether code does what it should, is secure, uses resources efficiently, and can be understood and changed over time.

What they found

The review found a substantial imbalance. Functional correctness has well-established evaluation frameworks and improvement strategies. Security increasingly does too. Performance efficiency and maintainability, however, remain significantly underexamined.

For those two less-studied dimensions, the authors report few standardized benchmarks and a lack of targeted fine-tuning approaches. This is a finding about gaps in the research landscape, not a measurement showing that every AI coding tool produces inefficient or hard-to-maintain code.

The authors propose several research directions: formal benchmarks, reinforcement learning that uses static and dynamic code feedback, and multi-agent systems that repeatedly critique and improve code using verifiable diagnostic evidence. These are proposed directions, not demonstrated improvements reported in the abstract.

Why it matters for builders

For someone building an app with AI coding tools, the practical takeaway is to avoid treating correctness as a complete definition of quality. A feature working as requested does not, by itself, answer questions about its security, efficiency, or ease of future modification.

Use the review’s four dimensions as separate questions when assessing generated code. Beyond checking behavior, consider what evidence you have about security, resource use, and maintainability. The review does not establish which coding assistant is best or validate a particular workflow. It highlights why confidence in one quality dimension should not automatically become confidence in all four.

Source: "An overview of evaluation and enhancement methods for code generation by large language models" (2026-05-12, Information and Software Technology). This is our summary of the paper's abstract — read the original for full details.