Multi-Hunk-APR

MultiFixer is a Coordinator-Proposer based multi-agent framework for fixing multi-hunk bugs.

Overview

Workflow

Defect Cause Analysis (Phase 1)

• Repository-level static analysis
  Conducted on both source code and test classes, with fine-grained inspection at the class, method, and class variable levels using static analysis tools (e.g., JavaParser).

• Agent + Tool interaction
  The BugAnalyzer agent autonomously investigates the root cause of the bug by iteratively invoking domain-specific tools (e.g., retrieve method body, get class hierarchy) based on analysis results. This mimics how developers navigate code in an IDE.

---

Repair Context Construction (Phase 2)

The repair context is composed of five key components:

• File/class/line-level contextual code
  Surrounding code at multiple granularities to provide dependency and control-flow information.

• Buggy hunk code
  The exact location of erroneous code, serving as the primary repair target.

• Failing test case code
  Relevant test cases that expose the defect, used to guide fix generation.

• Test failure report
  Error messages and stack traces from test execution, providing diagnostic clues.

• Bug analysis output (from Phase 1)
  Includes root cause summary and relevant code snippets identified by BugAnalyzer, enriching the semantic understanding of the bug.

Together, these form a comprehensive, multi-granular repair context.

---

Repair Process Based on the Proposer-Coordinator Architecture (Phase 3)

• Proposer
  A group of agents responsible for generating candidate patches for a given hunk. Multiple proposers run in parallel with diverse configurations (e.g., temperature, model variants) to enhance patch diversity.

• Coordinator

  • Selects the next hunk to repair based on dependencies and prior repair outcomes.
  • Evaluates all candidate patches generated by the Proposers and selects the most promising one.

• Hunk-level iterative repair workflow

  1. The Coordinator selects the next hunk to repair.
  2. All Proposers generate patch candidates for the selected hunk.
  3. The Coordinator evaluates and selects the optimal patch.
  4. Repeat until all hunks are fixed or the maximum iteration limit is reached.

This "propose-then-select" paradigm leverages the generation-recognition asymmetry to improve repair accuracy.

---

Iterative Patch Refinement (Phase 4)

After patch generation, two-stage refinement ensures correctness:

• Syntax Refinement
  If the patch fails to compile, the model uses compiler error messages to iteratively correct syntactic issues (e.g., missing braces, incorrect indentation).

• Test Refinement
  If tests still fail after successful compilation, the model analyzes test failure reports and refines the patch semantically (e.g., fixing logic errors).

Refinement proceeds iteratively until the patch passes all tests or reaches the maximum number of iterations.

Experiment

1. Perform static analysis on the buggy project.

java -jar artifacts/FileParser-1.0-SNAPSHOT-jar-with-dependencies.jar $PROJECT_DIR

2. Bug analysis

# for multi-hunk bugs
bash script/multi_hunk_bug_analysis.sh 
# for single-hunk bugs
bash script/single_hunk_bug_analysis.sh 

# for vulnerabilities
bash script/vul_analysis.sh

3. Repair bugs

# for multi-hunk bugs
bash script/multi_hunk_repair.sh
# for single-hunk bugs
bash script/single_hunk_repair.sh 

# for vulnerabilities
bash script/vul_repair.sh

Results

Overall performance of MultiFixer on Defects4J

Method	Patch Size	D4J-v1.2: SL	SH	SM	MM	SF	MF	PF	D4J-v2.0: SL	SH	SM	MM	SF	MF	PF	Total: CF	PF
ThinkRepair	≤125	52	78	98	0	98	0	-	47	81	107	0	107	0	-	205	-
ChatRepair	≤500	57	79	114	0	114	0	-	48	48	48	0	48	0	-	162	-
RepairAgent	117	52	67	86	4	88	2	96	48	61	71	3	73	1	90	164	186
PReMM	15	53	70	121	26	140	7	184	52	129	141	19	152	8	191	307	375
MultiFixer	≤30	49	74	120	28	141	7	182	56	98	144	34	158	20	230	326	412

Abbreviations:

• SL: Single Line, SH: Single Hunk, SM: Single Method
• MM: Multiple Methods, SF: Single File, MF: Multiple Files
• CF: Correct Fix (Total), PF: Plausible Fix

Overall performance of MultiFixer on VUL4J

Method	MultiFixer	FSV-Codex	FSV-finetuned	NTR	VRPILOT	APR4Vul	ChatRepair
CF (RCR)	24 (30.37%)	10.9 (13.79%)	9 (11.39%)	14 (17.72%)	14 (17.72%)	16 (20.25%)	15 (18.98%)

Abbreviations:

• CF: Correct Fix
• RCR: Repair Success Rate = CF / Total Bugs (79)