FeatX: Editing Software by Editing Features for Repository-Level Code Evolution

2026-07-01 · Source: cs.SE updates on arXiv.org · Field: Technology & Digital — Software Development & Engineering, Artificial Intelligence & Machine Learning · Depth: Expert, long

Summary

FeatX is a feature-oriented tool designed for LLM-assisted software evolution, addressing the limitations of code-centric paradigms that require manual context management. Released in 2026, FeatX extracts a hierarchical epic-feature structure from existing repositories with explicit feature-to-code mappings. It then employs a three-stage Evolution Agent—contextual expansion, localization & planning, and concrete code modification—to translate feature edits into code patches. The workflow is presented through four coordinated panels: Feature, CodeMap, Agent, and Diff. Evaluation across a user study and replay experiments on 38 real-world feature-editing commits demonstrated that FeatX significantly reduces NASA-TLX cognitive load by 41% (from 12.5 to 7.4) and improves SUS usability by 15% (from 73 to 84) compared to vanilla ChatGPT. It also achieved a 42.6% relative improvement in function-level modification localization F1 (0.385) over strong LLM baselines like Claude-opus-4.5 (0.270), at a substantially lower cost of \$0.07.

Key takeaway

For AI Engineers and Software Engineers managing complex codebases, FeatX demonstrates a superior approach to LLM-assisted software evolution. If you are struggling with high cognitive load and inaccurate code modifications using current LLM tools, consider adopting feature-oriented paradigms. This method significantly reduces development effort and improves modification localization accuracy, offering a cost-effective alternative to traditional code-centric LLM interactions. Explore tools that provide explicit feature-to-code mappings and multi-stage evolution agents.

Key insights

Feature-oriented LLM tools reduce cognitive load and improve localization accuracy for repository-level code evolution.

Principles

• Explicit feature-to-code mapping improves localization.
• Hierarchical feature abstraction simplifies complex repositories.
• Multi-stage agents enhance LLM-driven code modification.

Method

FeatX extracts hierarchical features, then a three-stage Evolution Agent (contextual expansion, localization & planning, code modification) translates feature edits into code patches.

In practice

• Use feature-level abstraction for LLM-assisted refactoring.
• Implement multi-panel UIs for complex agent workflows.
• Prioritize cost-efficient LLMs for development tools.

Topics

• Software Evolution
• Large Language Models
• Feature Engineering
• Code Refactoring
• Developer Tools
• Cognitive Load Reduction

Code references

• a496263365/FeatX
• Naccl/NBlog

Best for: Machine Learning Engineer, Research Scientist, AI Scientist, Software Engineer, AI Engineer

Related on AIssential

• Fragments: May 14 — LLMs are transforming software development by enabling rapid code porting and data transformation, but require careful integration to maintain predictability and foster human learning.
• SWE-InfraBench: Evaluating Language Models on Cloud Infrastructure Code — LLMs struggle with imperative IaC code modification, but iterative feedback and RAG significantly improve performance.
• Evaluating LLMs on Java Code Snippet Adaptation Using a Mutation-Injection Framework — A mutation-injection framework enables systematic, instruction-free evaluation of LLMs on Java code snippet adaptation.
• Alternatives for the EDIT tool of LLM agents — Token-efficient LLM agent editing can be achieved via line-specific checksum tags, improving reliability and speed.
• Feature Engineering with LLMs: Techniques & Python Examples — LLMs automate semantic feature engineering, transforming raw data into context-rich representations for improved ML performance.
• Where Did the Variability Go? From Vibe Coding to Product Lines by Regeneration — Vibe coding shifts software variability resolution to generation time, necessitating specification-driven regeneration for product lines.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.SE updates on arXiv.org.