Learning Project-wise Subsequent Code Edits via Interleaving Neural-based Induction and Tool-based Deduction
New research tackles the 'cross-file edit' problem, combining LLMs with refactoring tools for 40% better accuracy.
A team of researchers from institutions including Zhejiang University and the National University of Singapore has introduced TRACE (Tool-integrated RecommendAtion for Code Editing), a novel AI system designed to solve a core problem in AI-assisted software engineering. Current tools like Cursor excel at local, single-file edits but struggle with accuracy on cross-file changes, while broader project-wide solutions like CoEdPilot face efficiency bottlenecks. TRACE's key innovation is its hybrid architecture, which intelligently decides when to use a neural model for semantic reasoning and when to invoke a built-in IDE tool (like a rename refactor or a linter) for syntactic, rule-based edits.
The technical paper, accepted at the prestigious ASE 2025 conference, details two main challenges the team solved. First, they trained a neural model to act as a 'controller,' learning when to switch between the inductive neural module and the deductive tool module. Second, they developed a novel, fine-grained code editing representation to boost the neural model's prediction performance. This approach allows TRACE to maintain high accuracy across an entire project's scope without sacrificing the speed that developers demand, effectively pushing the boundary on the classic trade-off triangle of scope, accuracy, and efficiency in AI coding assistants.
- Hybrid Architecture: TRACE interleaves neural models for semantic edits with IDE tools (e.g., linters, refactors) for syntactic edits, unlike pure LLM approaches.
- Solves Cross-File Problem: Specifically designed to improve upon tools like Cursor, which see performance drops on edits that span multiple files.
- Learned Controller: Uses a neural model to decide when to invoke a tool versus when to use neural prediction, optimizing the workflow.
Why It Matters
This research could lead to the next generation of AI coding assistants that are truly effective for large-scale refactoring and feature work, not just single-line completions.