Recoleta
GitHub
Source note

Learning Project-wise Subsequent Code Edits via Interleaving Neural-based Induction and Tool-based Deduction

Authors
Chenyan Liu, Yun Lin, Yuhuan Huang, Jiaxin Chang, Binhang Qi, Bo Jiang, et al.
Source
arXiv
Published
Collection
Software Intelligence
Project Wise Code EditingCode IntelligenceCross File RefactoringLLM Tool IntegrationInteractive Code Assistance
Open arXivMarkdown source

Summary

TRACE is a project-wide code editing system that mixes LLM predictions with IDE tools such as rename and def-use analysis. It targets cross-file follow-up edits, where pure neural editors miss locations or cost too much to scan the whole project.

Problem

  • Developers often make incremental, project-wide edits such as refactors, bug fixes, and feature changes; the paper cites prior evidence that these account for over 70% of commit-history changes.
  • Existing tools trade off scope, accuracy, and speed: local editors work well in a small area, while project-wide neural systems pay high location cost and can miss or hallucinate cross-file edits.
  • Git-diff style edit labels are too coarse for training because one hunk can contain multiple edit semantics; the authors measure this in 18.04% of hunks.

Approach

  • TRACE predicts the next edit from the project, prior edits, and an optional prompt.
  • It interleaves neural induction for semantic edits with tool deduction for syntactic edit patterns. In simple terms: if the recent edit looks like a rename, signature update, clone update, or diagnose-fix pattern, TRACE calls IDE/LSP tools to find related edits; otherwise it runs neural location and generation models over code windows.
  • The system has three parts: an edit-composition invoker that decides when to call tools, an edit locator that labels lines and gaps as edit targets, and an edit generator that writes the code change.
  • The paper also adds a finer edit representation with 6 labels instead of the usual 3: <KEEP>, <REPLACE>, <DELETE>, <NULL>, <INSERT>, and <BLOCK-SPLIT>. This separates mixed edit semantics inside one hunk.
  • The representation is built by parsing code with Tree-sitter, aligning old and new tokens with LCS, then assigning line and inter-line edit labels.

Results

  • Evaluation covers 38K commits, 678 projects, and 5 programming languages.
  • Against prior systems such as CoEdPilot, GrACE, and CCT5, TRACE improves edit-location precision by 43.76%, recall by 9.96%, and edit-generation accuracy by 11.16%.
  • The edit-composition invoker reaches 92.45% precision and 94.63% recall for deciding tool invocation.
  • The new edit representation improves the neural edit locator by 14.57% and the edit generator by 7.40%.
  • In interactive edit simulation, TRACE reduces time cost by 14.40% and reaches 27.71% suggestion acceptance; the abstract also states an acceptance rate 6.15% higher than Cursor.
  • The paper reports a user study with 24 participants across 3 tasks and says TRACE leads on cross-file global edits, but the excerpt does not provide detailed task-level numeric scores.

Link