Week by week.
LLMs for Code — Foundations & Evaluation
Tokenization, pretraining incl. fill-in-the-middle, the model landscape, pass@k, and contamination-aware evaluation.
Autonomous SE Agents — & SWE-bench
The ReAct loop, agent anatomy and the Agent–Computer Interface, SWE-bench (Verified/Live/Pro), and reading results honestly.
Automated Program Repair
Localize → generate → validate; GenProg / SemFix / Angelix; the overfitting problem; neural and LLM-based repair.
Neural Code Representations
Tokens vs AST/graph; CodeBERT, GraphCodeBERT, CodeT5, UniXcoder; contrastive pretraining and code retrieval.
LLM-based Test Generation
The oracle problem; search-based (EvoSuite) to LLM-based (TestPilot, CoverUp, TestGen-LLM); coverage vs mutation.
Code Review, Summarization & Documentation
Code → words a developer trusts: automated review (CodeReviewer), summarization, documentation, and the metric problem (BLEU vs usefulness).
Vulnerability Detection & Secure Code
Learning-based detection (Devign/LineVul), the lab-to-reality gap, secure code generation ("Asleep at the Keyboard"), and hybrid analysis + LLM.