Research Activities

SELAB studies multi-agent and agentic AI systems for software engineering, focusing on code agents that detect vulnerabilities, localize bugs, review code, generate tests, repair programs, and document software changes.

Research Timeline

2025 ~ Present

CodeOrchestra: Multi-Agent & Agentic SE Pipeline

Agentic Code Agents, Agentic Reliability Evaluation, Multi-Agent Bug Report Generation, Agentic Bug Localization, Multi-Agent Review-to-Test, Multi-Agent Review-to-Repair, Multi-Agent Program Repair, Multi-Agent Commit Message Generation, Multi-Agent Vulnerability Detection

CodeOrchestra Agentic Code Agents Agentic Reliability Evaluation Multi-Agent Bug Report Generation Agentic Bug Localization Multi-Agent Review-to-Test Multi-Agent Review-to-Repair Multi-Agent Program Repair Multi-Agent Commit Message Generation Multi-Agent Vulnerability Detection
2024

Foundations for Agentic SE

Retrieval Agents, Structure-Aware Repository Agents, Security Bug Analysis Agents, Duplicate Detection Agents

RAG-LLM Agents AST Retrieval Agents Security Bug Agents Feature Selection Agents Duplicate Detection Agents sLLM Workflow Agents
2023

Applied AI & Software Mining

Bug Severity Prediction, Medical Data Mining, Emotion AI

CodeBERT APR Clinical NLP Bug Triage Voice Phishing Detection Emotion AI MOU Samsung Hospital
2021 ~ 2022

Deep Learning for SE

Bug Triage, Duplicate Detection, Program Repair

CNN-LSTM Triage BERT Duplicate SeqGAN Repair Bug Localization Agents
~ 2020

Foundation Research

Bug Report Mining, Developer Recommendation, Severity Prediction

Topic Modeling Social Network Analysis GAN-based Repair ACM SAC

Key Research Themes Multi-Agent First

Agentic SE - Flagship

CodeOrchestra

End-to-End Multi-Agent SE Automation

The umbrella architecture for SELAB research: orchestrator, planner, retriever, reviewer, repair, and documentation agents cooperate across vulnerability detection, bug reporting, localization, review-to-test, review-to-repair, program repair, and commit documentation. The research focus is reliable task decomposition, tool use, shared repository memory, evidence agents, and human-in-the-loop control for repository-scale software maintenance.

Orchestrator Agents Memory Agents Evidence Agents
Agentic Core - Flagship

CodeAgents

Agentic Code Agents for Repository Reasoning

Builds the agentic foundation used by every theme: planner agents decompose tasks, retriever agents gather repository evidence, executor agents run tools, memory agents preserve context, and reviewer agents verify outcomes. The objective is reproducible LLM-based software engineering through explicit agent roles, repository memory, verification loops, and measurable decision traces.

Planner Agents Retriever Agents Reviewer Agents
Agentic Reliability

EvalGuard

Threshold-Aware Evaluation for Agentic Software Engineering

Audits the reliability of agentic software-engineering decisions by checking whether LLM scores, verbalizer probabilities, and classifier outputs are safely converted into binary decisions. Starting from cross-artifact SATD evaluation, EvalGuard diagnoses score compression, threshold mismatch, AUROC-Recall divergence, and source-validation threshold recovery. It acts as a cross-cutting reliability layer for vulnerability detection, bug localization, review-to-test, review-to-repair, program repair, and commit documentation agents.

Score-Audit Agents Threshold-Diagnostic Agents Calibration Agents
Agentic Reporting

AgentReport

Multi-Agent Bug Report Generation

Coordinates symptom, trace, log, reproduction, and summarization agents to generate structured bug reports. The output captures expected and observed behavior, reproduction steps, environment details, and diagnostic evidence for localization and repair agents.

Symptom Agents Trace Agents Report Agents
Agentic Localization

AgentLocalization

Agentic Bug Localization

Coordinates report-analysis, trace-analysis, retrieval, ranking, and evidence-linking agents to localize buggy files, methods, and statements. These localization agents connect natural-language symptoms, execution traces, stack traces, and AST-level code structure into verifiable evidence for repair and review agents.

Retrieval Agents Suspiciousness-Ranking Agents Evidence-Linking Agents
Agentic Review

Review2Test

Multi-Agent Code Review-to-Test

Coordinates review-understanding, test-planning, test-generation, and execution agents to transform reviewer comments and pull request discussions into executable regression tests. Test results provide objective feedback for downstream repair and validation agents.

Review Agents Test Planner Agents Execution Agents
Agentic Review

Review2Repair

Multi-Agent Code Review-to-Repair

Turns review feedback into concrete repair plans through review-intent, patch-planning, repair, and validation agents. The agents preserve the reviewer intent while limiting unsupported or over-broad code changes through explicit evidence and test feedback.

Review-Intent Agents Patch-Planning Agents Validation Agents
Agentic Repair - Flagship

AgentRepair

Multi-Agent Program Repair

Coordinates localization, retrieval, patch-generation, test, and reviewer agents for automated program repair. The system emphasizes repository-aware context selection, test-guided patch ranking, and safeguards against plausible but incorrect fixes.

Repair Agents Patch-Ranking Agents Test Agents
Agentic Documentation

CommitChrono

Multi-Agent Commit Message Generation

Uses diff-analysis, issue-linking, convention-checking, and summary agents to explain what changed, why it changed, and how it relates to prior development history. Commit documentation becomes the final traceable step of the multi-agent maintenance pipeline.

Diff Agents Convention Agents Summary Agents
Agentic Security

SecuFlow

Multi-Agent Vulnerability Detection

Coordinates scanner, classifier, evidence, and security-review agents to detect vulnerability-prone code and security bug reports. Each finding is tied to code evidence, risk rationale, and downstream repair or test actions so vulnerability detection becomes part of the same agentic maintenance loop.

Scanner Agents Risk Agents Security-Review Agents