Winter 2027 · University of Alberta
CMPUT 663: AI for Software Engineering
- Term
- Winter 2027
- Instructor
- Prof. Zhou Yang
- Department
- Computing Science
Course overview
This graduate course explores how AI is built and used for software engineering, with a focus on large language models of code and coding agents. It connects these systems to software engineering tasks and examines how to evaluate their usefulness, reliability, and limitations.
The course brings together software engineering and machine learning, including the security, privacy, and human factors that matter when AI is used in real software development.
Interested in research? For University of Alberta students interested in doing research with us, we strongly recommend taking this course.
Course timeline
15 teaching weeks · 2 classes per week · Reading Week excluded
-
Week 01
Introduction to Software Engineering
This week introduces the goals, processes, tasks, and challenges of software engineering, covering requirements engineering, design, implementation, testing, deployment, and maintenance. Students will become familiar with software representations, formal specifications, testing, fuzzing, program analysis, agile development, and modern development tools, providing a foundation for understanding the engineering problems that AI can help address.
-
Week 02
Research Methods in Software Engineering
This week introduces the research process in software engineering, including research questions and hypotheses, literature search and review, research design and evaluation, academic writing, and peer review. Students will learn to critically read existing work, assess research claims and supporting evidence, identify open problems, and develop an initial proposal for their group project.
-
Week 03
Mining Software Repositories and Software Analytics
This week explores how source code, version histories, issue reports, code reviews, and execution logs can provide insights into software development and support task automation. Topics include data collection, feature and label design, classification, regression, clustering, and deep learning. Students will learn to construct research datasets and examine how data quality, sampling bias, and data leakage affect their findings.
-
Week 04
Code Representation Learning and Pretraining
This week examines code representation learning and language model pretraining, covering tokenization, model architectures, training data, and pretraining objectives. Examples include code completion, masked language modeling, and fill-in-the-middle training. Students will explore how data and training objectives shape model capabilities, and learn to prepare training examples and conduct small-scale training experiments.
Check Reading List:19 papers
- code2vec: Learning Distributed Representations of Code
- CodeBERT: A Pre-Trained Model for Programming and Natural Languages
- CodeXGLUE: A Machine Learning Benchmark Dataset for Code Understanding and Generation
- Evaluating Large Language Models Trained on Code
- Efficient Training of Language Models to Fill in the Middle
- CodeGen: An Open Large Language Model for Code with Multi-Turn Program Synthesis
- CodeGen2: Lessons for Training LLMs on Programming and Natural Languages
- Textbooks Are All You Need
- Code Llama: Open Foundation Models for Code
- DeepSeek-Coder: When the Large Language Model Meets Programming—The Rise of Code Intelligence
- StarCoder 2 and The Stack v2: The Next Generation
- Granite Code Models: A Family of Open Foundation Models for Code Intelligence
- CodeGemma: Open Code Models Based on Gemma
- DeepSeek-Coder-V2: Breaking the Barrier of Closed-Source Models in Code Intelligence
- Qwen2.5-Coder Technical Report
- Seed-Coder: Let the Code Model Curate Data for Itself
- Mellum: Production-Grade in-IDE Contextual Code Completion with Multi-File Project Understanding
- Qwen3-Coder-Next Technical Report
- Mellum2 Technical Report
-
Week 05
Post-Training Code Models
This week examines how pretrained code models are adapted to follow instructions, perform software engineering tasks, and learn from feedback. Topics include supervised fine-tuning and instruction tuning, preference optimization, knowledge distillation, and reinforcement learning with methods such as GRPO. Students will learn to design training examples and objectives, compare post-training methods, and investigate how learning signals affect model behavior and performance.
Check Reading List:19 papers
- Training Language Models to Follow Instructions with Human Feedback
- Self-Instruct: Aligning Language Models with Self-Generated Instructions
- Scaling Instruction-Finetuned Language Models
- LoRA: Low-Rank Adaptation of Large Language Models
- Direct Preference Optimization: Your Language Model is Secretly a Reward Model
- DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Language Models
- DeepSeek-R1 Incentivizes Reasoning in LLMs through Reinforcement Learning
- CodeRL: Mastering Code Generation through Pretrained Models and Deep Reinforcement Learning
- WizardCoder: Empowering Code Large Language Models with Evol-Instruct
- Magicoder: Empowering Code Generation with OSS-Instruct
- OpenCodeInterpreter: Integrating Code Generation with Execution and Refinement
- Training Language Models to Self-Correct via Reinforcement Learning
- Training Software Engineering Agents and Verifiers with SWE-Gym
- SWE-RL: Advancing LLM Reasoning via Reinforcement Learning on Open Software Evolution
- SWE-smith: Scaling Data for Software Engineering Agents
- SERA: Soft-Verified Efficient Repository Agents
- Toward Training Superintelligent Software Agents through Self-Play SWE-RL
- LEGO-RL: Harness-Native Reinforcement Learning for Coding Agents
- Does Reinforcement Learning Really Incentivize Reasoning Capacity in LLMs Beyond the Base Model?
-
Week 06
Inference-Time Scaling
This week explores how additional computation at inference time can improve performance on software engineering tasks. Topics include sampling multiple candidates, reasoning, search, iterative refinement, execution feedback, and verification. Students will learn to distinguish candidate generation from candidate selection, compare strategies under matched computational budgets, and analyze trade-offs among task success, latency, and computational cost.
Check Reading List:20 papers
- S*: Test Time Scaling for Code Generation
- CodeMonkeys: Scaling Test-Time Compute for Software Engineering
- DARS: Dynamic Action Re-Sampling to Enhance Coding Agent Performance by Adaptive Tree Traversal
- Wider or Deeper? Scaling LLM Inference-Time Compute with Adaptive Branching Tree Search
- Satori-SWE: Evolutionary Test-Time Scaling for Sample-Efficient Software Engineering
- Dynamic Scaling of Unit Tests for Code Reward Modeling
- Learning to Generate Unit Tests for Automated Debugging
- R2E-Gym: Procedural Environment Generation and Hybrid Verifiers for Scaling Open-Weights SWE Agents
- CodeChemist: Test-Time Scaling for Low-Resource Code Generation via Functional Knowledge Transfer
- Code Generation by Differential Test Time Scaling
- GraphAHA: Graph-Based Adaptive Search with Heterogeneous Actions for Test-Time Code Generation
- TipCoder: Reinforcement Learning Boosted Test-time Instruction Proposer for Code Generation
- Verification Trap: Understanding Test-Time Selection Failures under False Premises in Code Generation
- Adaptive Generate-Rank-Verify: Inference-Time Search with Costly Verification
- Verifiers in the Loop: Decoding Time Verification for Code Translation
- An Iterative Test-and-Repair Framework for Competitive Code Generation
- Towards Functional Correctness of Code Models with Selective Generation
- A Theoretical Analysis of Test-Driven Code Generation
- ExPairT-LLM: Exact Learning for LLM Code Selection by Pairwise Queries
- Reason-Code: Reliable Code Generation via Test-Driven Monte Carlo Tree Search
-
Week 07
Coding Agents
This week introduces agents that combine language models, tool use, and environmental feedback to perform software engineering tasks. Topics include task decomposition, repository exploration, planning, memory, and multi-agent collaboration. Through examples of fault localization, code modification, testing, and repair, students will learn to build a basic agent loop, analyze execution traces, and compare different workflows.
Check Reading List:20 papers
- LocAgent: Graph-Guided LLM Agents for Code Localization
- RepoMaster: Autonomous Exploration and Understanding of GitHub Repositories for Complex Task Solving
- Multi-CoLoR: Context-Aware Localization and Reasoning across Multi-Language Codebases
- PatchPilot: A Cost-Efficient Software Engineering Agent with Early Attempts on Formal Verification
- Unified Software Engineering Agent as AI Software Engineer
- Trae Agent: An LLM-based Agent for Software Engineering with Test-time Scaling
- debug-gym: A Text-Based Environment for Interactive Debugging
- RepoAudit: An Autonomous LLM-Agent for Repository-Level Code Auditing
- RefAgent: A Multi-agent LLM-based Framework for Automatic Software Refactoring
- Multi-Agent Coordinated Rename Refactoring
- RefactorAssist: Agentic Refinement for Reliable Code Refactoring
- Multi-Agent LLM Collaboration for Unit Test Generation via Human-Testing-Inspired Workflows
- Archer: Towards Agentic Review for Compiler Optimizations
- Adversarial Review: Structured Disagreement for Grounded Agentic Code Review
- SWE-MeM: Learning Adaptive Memory Management for Long-Horizon Coding Agents
- MemTrace: State-Consistent Memory for Long-Horizon Coding Agents
- SE-Agent: Self-Evolution Trajectory Optimization in Multi-Step Reasoning with LLM-Based Agents
- Scaling Inference-Time Compute with Teams of Coding Agents
- ChatDev 2.0: A No-Code Multi-Agent Platform for Developing Everything
- DeepCode: Open Agentic Coding
-
Week 08
Agent Harnesses
This week examines agent harnesses, the software infrastructure that supports and controls agent execution. Topics include tool interfaces, execution environments, context management, state persistence, resource budgets, and failure handling and recovery. Students will learn how these choices affect the information, actions, and feedback available to an agent, and gain experience modifying harness components and evaluating their interaction with the model.
Check Reading List:20 papers
- The OpenHands Software Agent SDK: A Composable and Extensible Foundation for Production Agents
- AgentScope 1.0: A Developer-Centric Framework for Building Agentic Applications
- Configuring Agentic AI Coding Tools: An Exploratory Study
- Harness Engineering: Anatomy, Architecture, and Evolution of Coding Agents -- A Source-Code Study of Eleven Systems
- Code as Agent Harness
- LLM-as-Code: Agentic Programming for Agent Harness
- The Complexity Trap: Simple Observation Masking Is as Efficient as LLM Summarization for Agent Context Management
- An Empirical Study of Harness Design for Coding Agents
- Beyond the Model: Demystifying Harness Effects in Software Engineering Agents
- CTXAgent - Self-Controlled Context Management of LLM Agents
- ACM: Agentic Context Management for Long Horizon Tasks
- Code Isn't Memory: A Structural Codebase Index Inside a Coding Agent
- RefactorPlatform: An Open-Source Harness for Controlled Evaluation of Repository-Scale Refactoring Agents
- Agentic Harness Engineering: Observability-Driven Automatic Evolution of Coding-Agent Harnesses
- Meta-Harness: End-to-End Optimization of Model Harnesses
- LEGO-RL: Harness-Native Reinforcement Learning for Coding Agents
- Effective Context Engineering for AI Agents
- Effective Harnesses for Long-Running Agents
- Harness Design for Long-Running Application Development
- Harness Engineering: Leveraging Codex in an Agent-First World
-
Week 09
AI for Software Engineering in Industry
Industry practitioners will share their experience developing and deploying AI tools for software engineering. Discussions will examine how models and agents fit into development workflows, including code review and team collaboration, and how quality, cost, and operational constraints influence adoption. Students will use concrete cases to assess benefits and limitations, understand engineering bottlenecks, and identify research opportunities.
Check Reading List:22 papers
- ★ "My productivity is boosted, but ..." Demystifying Users' Perception on AI Coding Assistants
- The SPACE of AI: Real-World Lessons on AI’s Impact on Developers
- DORA 2025 State of AI-assisted Software Development Report
- On the Use of Agentic Coding: An Empirical Study of Pull Requests on GitHub
- Programming by Chat: A Large-Scale Behavioral Analysis of 11,579 Real-World AI-Assisted IDE Sessions
- From Delegation to Action Approval: An Exploratory Study of Developers' Practices in Collaborating with AI Coding Agents
- Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity
- We Are Changing Our Developer Productivity Experiment Design
- Measuring the Self-Reported Impact of Early-2026 AI on Technical Worker Productivity
- Writing Code vs. Shipping Code: Productivity Effects Across Generations of AI Coding Tools
- AI Writes Faster Than Humans Can Review: A Longitudinal Study of an Enterprise 2x Mandate
- Agentic Coding and Persistent Returns to Expertise
- How AI Impacts Skill Formation
- The Impact of AI Coding Assistants on Software Engineering: A Longitudinal Study
- Does Working with AI Agents Change How Developers Code, Test, and Review?
- Testing with AI Agents: An Empirical Study of Test Generation Frequency, Quality, and Coverage
- Debt Behind the AI Boom: A Large-Scale Empirical Study of AI-Generated Code in the Wild
- Engineering Pitfalls in AI Coding Tools: An Empirical Study of Bugs in Claude Code, Codex, and Gemini CLI
- How AI Is Transforming Work at Anthropic
- Harness Engineering: Leveraging Codex in an Agent-First World
- Scaling Long-Running Autonomous Coding
- Agent Swarms and the New Model Economics
-
Week 10
Benchmarking and Evaluation
This week examines how software engineering benchmarks are constructed and evaluated, including task selection, data splits, execution environments, and evaluation metrics. We will consider functional correctness, the runtime performance of generated software, and agent efficiency, alongside data leakage, test adequacy, variability, and the reliability of human and model-based evaluation. Students will learn to design reproducible comparisons and assess which conclusions the results support.
Check Reading List:22 papers
- Multi-SWE-bench: A Multilingual Benchmark for Issue Resolving
- SWE-PolyBench: A multi-language benchmark for repository level evaluation of coding agents
- SWE-bench Goes Live!
- SWE-rebench V2: Language-Agnostic SWE Task Collection at Scale
- SWE-Lancer: Can Frontier LLMs Earn $1 Million from Real-World Freelance Software Engineering?
- FeatureBench: Benchmarking Agentic Coding for Complex Feature Development
- APEX-SWE
- Terminal-Bench: Benchmarking Agents on Hard, Realistic Tasks in Command Line Interfaces
- SWE-Serve: Benchmarking Agentic Engineering For Production Inference Serving
- SlopCodeBench: Benchmarking How Coding Agents Degrade Over Long-Horizon Iterative Tasks
- Multi-step SWE: Can Coding Agents Keep Building on Their Own Code Over Long Horizons?
- SWE-Together: Evaluating Coding Agents in Interactive User Sessions
- SWE-Gate: Passing Functional Tests Is Not Enough for Software Engineering Agents
- Code Review Agent Benchmark
- SWE-fficiency: Can Language Models Optimize Real-World Repositories on Real Workloads?
- Are “Solved Issues” in SWE-bench Really Solved Correctly? An Empirical Study
- SWE-Bench Pro V2: A Cleaner, Harder-to-Game Coding Benchmark
- Well-Specified and Well-Tested? Assessing SWE-bench-Style Datasets with a Self-Evolving Agent
- ★ Beyond Text Matching: Towards Reference-Free Evaluation for Human-Oriented Binary Reverse Engineering
- Quantifying Memorization Advantage in Code LLMs
- Reflections on the Reproducibility of Commercial LLM Performance in Empirical Software Engineering Studies
- Quantifying Infrastructure Noise in Agentic Coding Evals
-
Week 11
Reliability, Security, and Privacy
This week examines the reliability, security, and privacy of coding agents and generated software, as well as fairness requirements in relevant applications. Topics include secure code generation, prompt injection, unsafe tool use, access control, runtime verification, and human oversight. Students will learn to identify risks, evaluate mitigation strategies, and understand their guarantees, limitations, and effects on task completion.
Check Reading List:26 papers
- Teaching Code Agents to Ask Without Leaking Private Code
- PrivCode++: Latent-Conditioned Differentially Private Code Generation for Comprehensive Guarantees
- Scrub It Out! Erasing Sensitive Memorization in Code Language Models via Machine Unlearning
- PrivCode: When Code Generation Meets Differential Privacy
- Watermarking LLM Agent Trajectories
- Backdoors in Code Summarizers: How Bad Is It?
- Insecure Coding Preferences in Long-Term Memory: Security Risks for LLM-based Code Generation
- SkillJect: Effectively Automating Skill-Based Prompt Injection for Skill-Enabled Agents
- MCPTox: A Benchmark for Tool Poisoning on Real-World MCP Servers
- BaxBench: Can LLMs Generate Correct and Secure Backends?
- Is Vibe Coding Safe? Benchmarking Vulnerability of Agent-Generated Code in Real-World Tasks
- SecureVibeBench: Benchmarking Secure Vibe Coding of AI Agents via Reconstructing Vulnerability-Introducing Scenarios
- Benchmarking Correctness and Security in Multi-Turn Code Generation
- DeepGuard: Secure Code Generation via Multi-Layer Semantic Aggregation
- Learning from Failures: Error Notebook-guided Secure Code Generation
- AgentSpec: Customizable Runtime Enforcement for Safe and Reliable LLM Agents
- Defeating Prompt Injections by Design
- Safety Sidecar: Reflection-Driven Runtime Control for Safer Agents
- HarnessSecurity-Bench: Do Security Mechanisms Really Protect Coding Agent Harnesses?
- SABER: Benchmarking Operational Safety of LLM Coding Agents in Stateful Project Workspaces
- AttriGuard: Defeating Indirect Prompt Injection in LLM Agents via Causal Attribution of Tool Invocations
- The Attacker Moves Second: Stronger Adaptive Attacks Bypass Defenses Against LLM Jailbreaks and Prompt Injections
- How Reasoning Shapes Social Bias in LLM-Generated Code?
- FairCoder: Probing LLM Bias in High-Stakes Decision Making via Coding Tasks
- From If-Statements to ML Pipelines: Revisiting Bias in Code-Generation
- Hidden Licensing Risks in the PTMware Ecosystem
-
Week 12
Applications in Hardware, Robotics, and Algorithm Optimization
This week explores applications of code models and agents in hardware and integrated circuit design, robotics, and algorithm optimization. Through examples of domain adaptation, code generation, tool use, and verification, students will compare the domain knowledge, program representations, execution environments, and feedback required in each setting. They will examine where existing methods transfer and identify new engineering challenges.
Check Reading List:30 papers
- Comprehensive Verilog Design Problems: A Next-Generation Benchmark Dataset for Evaluating Large Language Models and Agents on RTL Design and Verification
- FormalRTL: Verified RTL Synthesis at Scale
- Veri-Sure: Multi-Agent RTL Code Generation with Temporal Tracing, Slicing and Formal Verification
- SYMDIREC: A Neuro-Symbolic Divide-Retrieve-Conquer Framework for Enhanced RTL Synthesis and Summarization
- Exploring the Agentic Frontier of Verilog Code Generation
- GoGoTB: Agentic RTL Verification with Specification-Grounded Coverage Closure
- Closing the Loop on LLM-Generated RTL Assertions with Quality-Aware Formal Verification
- Will AI mess my RTL? A snapshot-based approach to verifying AI-written RTL
- LongRTL: Graph-Similarity-Guided LLM-driven Long Context RTL Optimization
- Toward Autonomous RTL and Synthesis Optimization: Benchmarking and Self-Improving Agents
- VPR-Evolve: Multi-Agent-Driven Algorithm Evolution for FPGA Place and Route
- AlphaEvolve: A Coding Agent for Scientific and Algorithmic Discovery
- Kevin: Multi-Turn RL for Generating CUDA Kernels
- CUDA-L1: Improving CUDA Optimization via Contrastive Reinforcement Learning
- Geak: Introducing Triton Kernel AI Agent & Evaluation Benchmarks
- KernelFalcon: Autonomous GPU Kernel Generation via Deep Agents
- CUDA Agent: Large-Scale Agentic RL for High-Performance CUDA Kernel Generation
- KernelEvolve: Scaling Agentic Kernel Coding for Heterogeneous AI Accelerators at Meta
- KernelOPT: Dispatch-Aware Agentic Search for GPU Kernel Optimization
- Are LLM-Generated GPU Kernels Production-Ready? A Trace-Driven Benchmark and Optimization Agent
- HIERA: Workload-Aware Planning Across Implementation Spaces for GPU Kernel Optimization
- AutoMisty: A Multi-Agent LLM Framework for Automated Code Generation in the Misty Social Robot
- CaP-X: A Framework for Benchmarking and Improving Coding Agents for Robot Manipulation
- RHO: Your Coding Agent is Secretly a Roboticist
- ENPIRE: Agentic Robot Policy Self-Improvement in the Real World
- ASPIRE: Agentic /Skills Discovery for Robotics
- SEAR: Self-Evolving Agentic Robotics
- RoboAuto: Autonomous Robot Program Search through Multi-Level Physical Understanding
- ACE-CaP-0: Real-World Code-as-Policies in the Era of Experience
- ★ ManimAgent: Self-Evolving Multimodal Agents for Visual Education
-
Week 13
Research Project Presentations I
The first set of groups will present their research projects and papers, covering the research questions, related work, methods, experimental results, and limitations. Through demonstrations, questions, and peer feedback, students will explain how the evidence supports their contributions, discuss reproducibility, and identify revisions needed to strengthen their work.
-
Week 14
Research Project Presentations II
The remaining groups will present their research projects and papers, covering the research questions, related work, methods, experimental results, and limitations. Through demonstrations, questions, and peer feedback, students will explain how the evidence supports their contributions, discuss reproducibility, and identify revisions needed to strengthen their work.
-
Week 15
Project Revisions and Research Artifacts
Groups will present their revisions and respond to feedback from earlier sessions. Discussions will focus on additional evidence, changes to methods or conclusions, remaining limitations, and reproducibility. Students will refine their papers and research artifacts, justify their final claims, and identify the work needed before submission to a research venue.
Course materials
The weekly topics and available reading lists are included in the timeline above. More readings and lecture materials will be added here as they become available.
- Course timeline
- View weekly topics
- Reading lists
- Week 04 · 19 papers
Week 05 · 19 papers
Week 06 · 20 papers
Week 07 · 20 papers
Week 08 · 20 papers
Week 09 · 22 papers
Week 10 · 22 papers
Week 11 · 26 papers
Week 12 · 30 papers - Lecture materials
- To be added
Contact
For course enquiries, email zy25@ualberta.ca.