Yiling Lou

Course Introduction

Artificial intelligence is reshaping modern software systems, with generative AI models increasingly integrated to provide intelligent functionality. This course explores emerging software engineering methods and practices for designing, developing, and operating AI-integrated software. Topics include recent advances in software quality assurance and maintenance techniques (such as testing, debugging, security, efficiency optimization, and DevOps) for agentic systems. We will also discuss testing and debugging techniques for AI libraries and compilers, as well as the broader ecosystem supporting AI-integrated software.

This course aims to offer an overall picture of software engineering for generative AI. It will be in seminar format, focusing on cutting-edge research papers and semester-long projects. It covers a wide range of topics including (1) basis for SE and GenAI, (2) quality assurance for agents, (3) quality assurance for AI infrastructure, (4) quality assurance for domain-specific AI-powered software, (5) non-functional property in AI-powered software. The goal of this course is to get students familiar with the SE-for-AI area and also the research process.

Instructor: Yiling Lou (yilingl@illinois.edu)
Class Time: Mon/Wed 09:30AM - 10:45AM (Central Time)
Office Hours: Mon/Wed 10:45AM - 11:45AM (Central Time)
Location: Siebel Center for Comp Sci 0218
Forum/project Submissions:: Campuswire (Create an account with netid@illinois.edu email address, you should already have access; join before your first class as all notifications will be there; ask questions on Campuswire and email the instructor for private issues.)

Grading

Class Participation (15%): The course is centering on discussion. This item would be evaluated on the class attendance and discussion participation (e.g., ask or answer questions). Please read at least the primary papers before each class to ensure engagement in class.
Paper Presentation (25%): Each student is required to present one paper and lead its discussion. (1) Register the preference for topics and papers (select at least five topics) in Campuswire "Assignment" before Jan. 30th. (2) Prepare slides for presentation by yourself rather than directly using original author slides. Upload the initial version of slides to Campuswire a week before the presentation slot to get feedback comments. Upload the final version of slides to Campuswire 24h before the presentation slot. (3) Lead the discussion of the paper and answer questions from audience. Prepare some good questions to initiate the discussion.
Course Project (60%): The semester-long course project is based on groups (2-3 students per group). There is a list of candidate topics for reference in Campuswire. In addition, proposing your own topic is allowed, but meet with the instructor to discuss the proposal before Feb. 10th. There are three check points for the course project as follows. The deliverables of each stage include both the report and presentation. Detailed/deadlines for each phase are in the schedule section below.
- Proposal report and presentation (5%): 1-page .txt proposal (due 02/26) and 5 min presentation (02/23).
- Midterm report and presentation (20%): 3-page PDF report (due 04/13) and 10 min presentation (04/06, 04/08).
- Final report and presentation (35%): 5-page PDF report (due 05/13) and 12 min presentation (05/04, 05/06).

Tentative Schedule

Date	Topics	Notes
01/21 Wed	Course Introduction	Presenter: Yiling
01/26 Mon	SE Basis	Presenter: Yiling
01/28 Wed	SE Basis	Presenter: Yiling
02/02 Mon	SE Basis	Presenter: Yiling
02/04 Wed	GenAI Basis	Presenter: Yiling
02/09 Mon	LLM Agents: Memory Cognitive Architectures for Language Agents MemGPT: Towards LLMs as Operating Systems A-Mem: Agentic Memory for LLM Agents
02/11 Wed	LLM Agents: Action Gorilla: Large Language Model Connected with Massive APIs Toolformer: Language Models Can Teach Themselves to Use Tools Large Language Models as Tool Makers
02/16 Mon	LLM Agents: Multi-Agent System AutoGen: Enabling Next-Gen LLM Applications via Multi-Agent Conversation CAMEL: Communicative Agents for “Mind” Exploration of Large Language Model Society Towards a Science of Scaling Agent Systems
02/18 Wed	LLM Agents: Application SWE-agent: Agent-Computer Interfaces Enable Automated Software Engineering Agent Laboratory: Using LLM Agents as Research Assistants Agentless: Demystifying LLM-based Software Engineering Agents OpenHands: An Open Platform for AI Software Developers as Generalist Agents
02/23 Mon	Proposal Presentation
02/25 Wed	Understanding Bugs in Agents Taxonomy of Failure Modes in Agentic AI Systems Why Do Multi-Agent LLM Systems Fail? Can Agents Fix Agent Issues? Are LLMs Correctly Integrated into Software Systems?
03/02 Mon	Testing for Agents (1): Benchmark Autonomous Evaluation and Refinement of Digital Agents SEC-bench: Automated Benchmarking of LLM Agents on Real-World Software Security Tasks AgentBench: Evaluating LLMs as Agents SWE-bench: Can Language Models Resolve Real-World GitHub Issues? Survey on Evaluation of LLM-based Agents Establishing Best Practices for Building Rigorous Agentic Benchmarks
03/04 Wed	Testing for Agents (2): Vulnerability Detection Make Agent Defeat Agent: Automatic Detection of Taint-Style Vulnerabilities in LLM-based Agents Demystifying RCE Vulnerabilities in LLM-integrated Apps Security Debt in LLM Agent Applications: A Measurement Study of Vulnerabilities and Mitigation Trade-offs AGENTVIGIL: Generic Black-Box Red-teaming for Indirect Prompt Injection against LLM Agents
03/09 Mon	Debugging for Agents (1) Which Agent Causes Task Failures and When? On Automated Failure Attribution of LLM Multi-Agent Systems Process-Centric Analysis of Agentic Software Systems Where LLM Agents Fail and How They Can Learn From Failures
03/11 Wed	Debugging for Agents (2) Interactive Debugging and Steering of Multi-Agent AI Systems Dover: Intervention-driven Auto Debugging for LLM Multi-Agent Systems Comfrey: Mitigating Integration Failures in LLM-enabled Software at Run-Time
03/16 Mon	Spring Break
03/18 Wed	Spring Break
03/23 Mon	Agent Observability TraceAegis: Securing LLM-Based Agents via Hierarchical and Behavioral Anomaly Detection AgentSight: System-Level Observability for AI Agents Using eBPF A Survey on AgentOps: Categorization, Challenges, and Future Directions AI Agent Observability – Evolving Standards and Best Practices
03/25 Wed	Self-Improving Agents DSPY: Compiling Declarative Language Model Calls into Self-Improving Pipelines Godel Agent: A Self-Referential Agent Framework for Recursive Self-Improvement LIVE-SWE-AGENT: Can Software Engineering Agents Self-Evolve on the Fly? Darwin GÖDEL Machine: Open-Ended Evolution of Self-Improving Agents Multi-Agent Evolve: LLM Self-Improve Through Co-Evolution A Survey of Self-Evolving Agents: On the Path to Artificial Super Intelligence
03/30 Mon	Build/Deploy Agents in Practice Measuring Agents in Production An Empirical Study of Agent Developer Practices in AI Agent Frameworks
04/01 Wed	Testing & Debugging AI Program DeepFD: Automated Fault Diagnosis and Localization for Deep Learning Programs A Study on Prompt Design, Advantages and Limitations of ChatGPT for Deep Learning Program Repair
04/06 Mon	Midterm Presentation
04/08 Wed	Midterm Presentation
04/13 Mon	Testing & Debugging AI Libraries Large Language Models are Zero-Shot Fuzzers: Fuzzing Deep-Learning Libraries via Large Language Models Toward Understanding Deep Learning Framework Bugs Large Language Models are Edge-Case Fuzzers: Testing Deep Learning Libraries via FuzzGPT Deep Learning Library Testing: Definition, Methods and Challenges
04/15 Wed	Testing & Debugging AI Compilers NNSmith: Generating Diverse and Valid Test Cases for Deep Learning Compilers PolyJuice: Detecting Mis-compilation Bugs in Tensor Compilers with Equality Saturation Based Rewriting Coverage-Guided Tensor Compiler Fuzzing with Joint IR-Pass Mutation
04/20 Mon	Testing & Debugging AI Software: ADS MoDitector: Module-Directed Testing for Autonomous Driving Systems VioHawk: Detecting Traffic Violations of Autonomous Driving Systems through Criticality-Guided Simulation Testing DeepTest: Automated Testing of Deep-Neural-Network-driven Autonomous Cars
04/22 Wed	Testing & Debugging AI Software: Embodied AI SENTINEL: A Multi-Level Formal Framework for Safety Evaluation of LLM-based Embodied Agents Subtle Risks, Critical Failures: A Framework for Diagnosing Physical Safety of LLMs for Embodied Decision Making IS-Bench: Evaluating Interactive Safety of VLM-Driven Embodied Agents in Daily Household Tasks AGENTSAFE: Benchmarking the Safety of Embodied Agents on Hazardous Instructions
04/27 Mon	Non-functional Property in GenAI (1) Cut the Crap: An Economical Communication Pipeline for LLM-based Multi-Agent Systems Reasoning in Token Economies: Budget-Aware Evaluation of LLM Reasoning Strategies Metron: Holistic Performance Evaluation Framework for LLM Inference Systems Fast Inference for Augmented Large Language Models SWE-Effi: Re-Evaluating Software AI Agent System Effectiveness Under Resource Constraints
04/29 Wed	Non-functional Property in GenAI (2) The Good, The Bad, and The Greedy: Evaluation of LLMs Should Not Ignore Non-Determinism LLM is Like a Box of Chocolates: The Non-determinism of ChatGPT in Code Generation Benchmarking Bias in Large Language Models during Role-Playing FLEX: A Benchmark for Evaluating Robustness of Fairness in Large Language Models
05/04 Mon	Final Project Presentation
05/06 Wed	Final Project Presentation