When Code Repairs Go Too Far: The Hidden Costs of Over-Editing in Large Language Models

In the fast-evolving field of software engineering, large language models (LLMs) have emerged as powerful tools for aiding code editing and repairs. However, a recent study by researchers Tongyao Zhu, Wei Hern Lim, and Min-Yen Kan at the National University of Singapore highlights a serious concern: the phenomenon of over-editing. This occurs when models modify code far beyond what's necessary for a successful fix, leading to potential confusion and difficulties in code reviews.

The Dilemma of Edit Fidelity

While LLMs like GPT-5.5 can produce functionally correct code repairs, the researchers discovered that these models often introduce unnecessary changes, complicating the code's structure. Over-editing can add cognitive complexity, making it harder for developers to review changes and understand the original intent of the code. In their study, the team evaluated the performance of several LLMs on 400 coding tasks drawn from the BigCodeBench benchmark, revealing alarming trends in edit fidelity.

Assessing Over-Editing: A Rigorous Framework

To better understand over-editing, the researchers developed a controlled evaluation framework. They identified minimal patches required to fix specific coding bugs and compared model outputs against these standards. Results showed that many top LLMs achieve high functional success rates but often make excessive modifications, generating a wide range of code changes unrelated to the core issue.

For example, in an off-by-one error, a minimal fix required only one line to be changed, but models like GPT-5.4 inexplicably added dozens of lines of code, introducing unnecessary complexity and making the edits more difficult to review.

Mitigation Strategies: Instructions Matter

The research also explored ways to reduce over-editing. By including explicit instructions to preserve the original code, the team found a significant reduction in excessive edits. This simple change lowered the average excess edit distance across models and improved the quality of the code produced. Notably, prompting the models with preservation instructions reduced added cognitive complexity by over 26%, a promising finding for developers looking to harness the power of LLMs while minimizing the drawbacks of over-editing.

Learning from Experience: The Future of Minimal Editing

The study further posits that minimal editing practices can be learned through reinforcement learning techniques. By training models to prioritize smaller, more faithful repairs, developers could significantly enhance the code repair process without sacrificing accuracy. This approach suggests that over-editing is not merely a flaw in current models but rather a behavior that can be modified with better training methodologies.

The Importance of Edit Fidelity in Software Development

As software development increasingly relies on AI tools, the implications of over-editing could pose a significant risk to efficiency and code quality. The study emphasizes the necessity of evaluating not just whether a model can produce correct code, but also how well it can do so with minimal disruption. Harnessing insights from this research could lead to the development of more effective AI coding assistants that improve coding workflows, reduce the burden of code reviews, and ultimately produce cleaner, more maintainable code.

As we continue to explore the capabilities of large language models in software engineering, this research serves as a crucial reminder: sometimes less really is more when it comes to code edits.