SELAB RESEARCH
We study whether the evidence software engineering relies on — tests, gates, and benchmarks — still holds when the code is written by AI.
The repository-grounded agents and tools we build are the instruments for that study, not the goal in themselves.
Current Program
Four questions run in sequence: whether our measurements are valid, what risk survives them, what an attacker can leave behind, and what a pipeline can actually stop. A fifth asks whether any of it still holds as models change. These are open questions under active study, not settled results.
MEASUREMENT VALIDITY · IN PROGRESS
Whether an evaluation protocol measures agent capability at all, and whether protocol choice alone can reverse a published conclusion. Examined through repeated runs and protocol transfer.
RESIDUAL RISK
How much behavioral failure remains in AI-generated patches that pass every test, and under what conditions an acceptance test can count as reliability evidence.
AI SOFTWARE SECURITY · IN PROGRESS
Whether repository-level prompt injection ends at attack success or leaves durable unsafe artifacts in the codebase, and whether those artifacts pass an ordinary development pipeline.
SOFTWARE ASSURANCE
Whether tests, static analysis, and CI checks detect different defects or the same ones, and which combinations justify their cost inside an assurance pipeline.
TEMPORAL VALIDITY
Whether a conclusion obtained today survives a change of model generation, and how model drift bounds the useful lifetime of an empirical finding.
Systems We Build
These questions concern the systems we build to run the studies above: how an agent grounds itself in a repository, acts through tools, verifies its own output, and when coordination earns its complexity.
REPOSITORY GROUNDING
We connect code structure, history, issues, logs, traces, and tests while measuring when retrieved context is missing, noisy, or unsupported.
TOOL-BASED ACTION
We study task decomposition, bounded tool use, memory, and recovery across analyzers, build systems, tests, and version control.
VERIFICATION & RELIABILITY
We evaluate executable correctness, regressions, uncertainty, traceability, and failure detection instead of accepting fluent outputs as evidence.
COORDINATION & HUMAN CONTROL
We compare task success, cost, latency, and failure propagation so coordination is selected only when its measurable benefit justifies its complexity.
Durable Agenda
The axes organize reusable scientific questions; individual systems and papers contribute evidence within them rather than becoming disconnected topics.
GROUNDING
Grounding decisions in repository-scale evidence through structural retrieval, issue–code linking, execution traces, and context sufficiency analysis.
ACTION
Planning and executing software-engineering work through developer tools while preserving permissions, constraints, and an observable action trace.
VERIFICATION
Generating tests, repairs, security findings, and documentation, then validating them against behavior, regressions, and human intent.
GOVERNANCE
Measuring robustness, calibration, traceability, cost, and latency while comparing single-agent, multi-agent, and human-in-the-loop designs.
Current Work
Composable Agentic Workflows for Software Maintenance
An experimental architecture for composing planning, retrieval, execution, memory, and verification across repository-scale tasks. Single-agent or multi-agent coordination is selected only when role separation or independent checking is an explicit, testable design choice.
Repository-Grounded Software Agents
Studies how software agents decompose goals, retrieve repository evidence, operate development tools, preserve task context, and verify outcomes through measurable decision and execution traces.
Calibrated Decisions for Agentic Software Engineering
Examines how model scores and classifier outputs become decisions, diagnosing score compression, threshold mismatch, metric divergence, and transfer failures across software artifacts and tasks.
Evidence-Grounded Bug Reporting
Studies structured bug-report generation from symptoms, logs, traces, reproduction steps, and environment evidence so downstream localization and repair begin with auditable problem statements.
Repository-Aware Bug Localization
Connects natural-language symptoms, execution traces, stack traces, history, and code structure to rank suspicious files, methods, and statements with inspectable supporting evidence.
Review-Grounded Test Generation
Transforms reviewer comments and pull-request discussions into explicit test objectives and executable regression tests, using execution results as feedback for validation and repair.
Review-Guided Program Repair
Turns review intent into constrained repair plans and candidate patches, then tests whether changes address the feedback without introducing unsupported or over-broad modifications.
Verified Repository-Aware Program Repair
Studies localization, context retrieval, patch generation, ranking, and test-guided validation with safeguards against plausible but behaviorally incorrect fixes.
Evidence-Linked Change Documentation
Links diffs, issues, project conventions, and development history to explain what changed, why it changed, and which evidence supports the resulting documentation.
Evidence-Aware Vulnerability Analysis
Studies how analyzers and language models combine code evidence, risk rationale, and validation steps to reduce unsupported security findings and connect them to tests or repairs.
Reliability by Design
Every claim is tied to an explicit task, baseline, observable trace, and failure condition.
Define the task, dataset, constraints, and success condition before choosing an agent architecture.
Compare non-agentic, single-agent, and multi-agent designs under the same tools, budget, and evidence.
Validate claims with tests, traces, regressions, calibration, and explicit rejection when evidence is insufficient.
Report ablations, costs, latency, failed cases, recovery behavior, and the points where human review is required.
Research Evolution
Our current Agentic AI agenda builds on earlier work in software analytics, deep learning, repository mining, and program repair.
Ongoing study of measurement validity, residual failure after passing tests, persistent unsafe artifacts from coding agents, complementarity of assurance gates, and whether empirical conclusions survive model generations.
Repository-grounded agents, tool-using workflows, executable verification, reliability evaluation, and evidence-aware coordination for software maintenance.
Retrieval-augmented reasoning, structure-aware repository context, security-bug analysis, feature selection, and duplicate detection.
Program repair, bug severity and triage, clinical NLP, voice-phishing detection, and emotion-aware applications.
Deep learning for bug triage, duplicate detection, bug localization, and program repair.
Bug-report mining, developer recommendation, severity prediction, topic modeling, and social-network analysis.