How AI coding agents can keep making correct changes to an evolving codebase: what past evidence to reuse, when to check again, and how to recover.
Research Trajectory
Our current agenda, keeping AI-changed software correct as it evolves, builds on earlier work in software analytics, deep learning, repository mining, and program repair.
Sustained Software Change by AI Coding Agents
One primary track and two related tracks. They grow out of our studies of memory in AI coding agents and extend earlier work on evaluation validity, residual failures behind passing tests, and assurance gates.
PRIMARY TRACKCurrent Focus
Keeping Code Correct Across Successive Changes
As an agent extends code it wrote earlier through changed requirements, updated libraries, and handoffs, which past successes and verification records can it still rely on? What must it recheck within a limited budget, and when should it repair forward or partially roll back? We judge methods by correctly completed changes, preserved behavior, and total maintenance cost, not by memory accuracy alone.
Experience Reuse
Reverification
Recovery
Maintenance Cost
ADJACENT TRACK
Acquiring the Evidence a Correct Fix Needs
Before changing code, an agent can search documentation, locate call sites, write a reproduction, compare versions, add a test or ask the requester. We study which of these actions resolves the uncertainty that matters at the least cost, and how to tell a change that can be judged from the specification apart from one that needs the requester's intent. How code and evidence are presented to the agent is treated as an experimental factor.
Specification Sufficiency
Evidence Acquisition
Clarifying Questions
Context Presentation
EXPLORATORY TRACK
Independent Judgment Among Multiple Agents
A second AI reviewer helps only if its judgment is independent. We examine how shared review records change what a second reviewer actually adds, and, once real cases show the problem is common enough, how parallel agents can integrate changes that each pass their own checks but fail together.
Correlated Misses
Review Independence
Parallel Integration
Recent 3 Years (2023-2025)
LLM Agents and Deep Learning for Bug Management19 papersAgents · GenerationLocalizationTriage · SecurityCode · Systems
Agents · Generation5 papers
LLM Agents for Bug Reports and Repair
Multi-agent and retrieval-augmented LLMs that generate reproducible bug reports and repair programs, including reinforcement learning and chain-of-thought variants.
Multi-Agent LLM
Bug Report Generation
Program Repair
RAG
Localization4 papers
Structure-Aware Bug Localization
Zero-shot and learned bug localization using AST-aware retrieval, LSTM attention, and causal representation learning.
Bug Localization
AST Retrieval
Attention
Causal Learning
Triage · Security6 papers
Bug Triage, Priority, and Security Classification
Deep learning for developer assignment, priority, duplicates, and security bug reports, including cross-project data augmentation and a stop-word study.
Bug Triage
Security Bug Reports
Duplicate Detection
Data Augmentation
Code · Systems4 papers
Commit Messages, Test Generation, and Systems
Context-aware commit message generation, coverage-guided test generation, diffusion-based minimal repair, and etcd performance under Kubernetes resource limits.
Commit Messages
Test Generation
Diffusion Models
Kubernetes
Security Bug Reports and Duplicate Detection3 papersSecurityDuplicates
Security2 papers
Security Bug Report Prediction
Severity-based feature selection with LSTM, and RAG with small LLMs to generate and predict security bug report templates.
Security Bug Reports
RAG
Small LLMs
Feature Selection
Duplicates1 paper
Feature Selection for Duplicate Detection
A scenario-based analysis of how feature selection affects duplicate bug report detection.
Duplicate Detection
Feature Selection
Bug Report Mining, Program Repair, and Applied AI12 papersMiningLocalization · RepairApplied AI
Mining6 papers
Bug Report Mining
Topic- and feature-based models for severity, priority, duplicates, and developer recommendation, including reinforcement learning on structured report representations.
Severity Prediction
Topic Modeling
Duplicate Detection
Developer Recommendation
Localization · Repair3 papers
Similarity-Based Localization and Repair
Similar code and stack traces for bug localization and C program repair, with an analysis of how similar code affects both.
Bug Localization
Program Repair
Code Similarity
Applied AI3 papers
Applied Deep Learning
Voice-phishing detection, sleep-quality prediction from lifelogs, and multimodal emotion recognition from speech and text.
Voice Phishing
Lifelog
Multimodal Emotion
Others
Deep Learning for Bug Triage, Localization, and Repair6 papersBug ReportsLocalization · Repair
Bug Reports3 papers
Deep Learning on Bug Reports
CNN-LSTM and fine-tuned BERT models for bug triage, severity prediction, and duplicate detection.
CNN-LSTM
BERT
Bug Triage
Duplicate Detection
Localization · Repair3 papers
Learned Localization and Repair
Similar-commit CNN-LSTM bug localization, code-block LSTM repair, and CodeBERT for multi-chunk repair.
Bug Localization
Program Repair
CodeBERT
Software Analytics for Bug Management28 papersSeverityTriageRepairApplied AI
Severity9 papers
Bug Severity and Priority Prediction
Topic models, emotion words, concept profiles, and cross-project similarity for predicting bug severity and priority.
Severity Prediction
Topic Modeling
Cross-Project
Triage10 papers
Bug Triage and Developer Recommendation
Topic models, developer networks, and role analysis for assigning bugs to developers, plus duplicate bug report detection.
Bug Triage
Developer Networks
Duplicate Detection
Repair7 papers
Automated Program Repair
Genetic programming with similar fixes, SeqGAN- and GAN-based repair, and deep learning for localization and repair.
Genetic Programming
SeqGAN
Program Repair
Applied AI2 papers
Applied AI for Broadcasting
An election-result prediction application and a deep-learning interactive broadcast service system.