Recoleta
GitHub
Source note

ARISE: A Repository-level Graph Representation and Toolset for Agentic Fault Localization and Program Repair

Authors
Shahd Seddik, Fatemeh Fard
Source
arXiv
Published
Collection
Software Intelligence
Software Engineering AgentsAutomated Program RepairFault LocalizationProgram AnalysisCode Graphs
Open arXivMarkdown source

Summary

ARISE adds a repository graph with statement-level def-use edges to an LLM repair agent. On SWE-bench Lite, it improves fine-grained localization and raises Pass@1 over SWE-agent.

Problem

  • Repository-level repair agents must find the right files, functions, and lines before they can patch a bug.
  • Existing graph tools mostly track files, classes, functions, imports, calls, and references; they miss how variable values move inside a function.
  • This matters because line- and function-level localization are a main failure point in SWE-bench-style repair.

Approach

  • ARISE builds a typed graph for Python repositories with Directory, Module, Class, Function, Method, and Statement nodes.
  • It adds Contains, Imports, Calls, Inherits, and intra-procedural DataflowDefUse/DataflowUseDef edges.
  • The data-flow pass scans each function body, records top-level statements, finds variable definitions and uses, and connects each use to the last preceding definition in the same function.
  • The agent gets a three-tier API: structural navigation, data-flow slicing, and context bundling. The main tool returns backward, forward, or bidirectional slices for a chosen variable and statement.
  • The evaluation uses SWE-agent with Qwen2.5-Coder-32B-Instruct on SWE-bench Lite.

Results

  • SWE-bench Lite contains 300 real GitHub issues from 11 Python repositories.
  • Against unmodified SWE-agent, ARISE improves Function Recall@1 by 17.0 points.
  • Against the same baseline, ARISE improves Line Recall@1 by 15.0 points.
  • ARISE reaches 22.0% Pass@1, fixing 66/300 issues, a 4.7 percentage-point gain over SWE-agent. This implies the baseline fixed about 52/300 issues, or 17.3% Pass@1.
  • Ablations attribute the gain to the data-flow graph; the tool-schema-only condition does not match it.
  • The paper reports that Qwen2.5-Coder-32B-Instruct can use structured slice output directly, without a natural-language summary layer.

Link