CS 6158: Software Engineering in the Era of AI (Cornell, Fall 2026)
by Cornell University (Saikat Dutta)
A research-paper reading list on coding agents, LLM-driven verification, fuzzing, code translation and testing ML systems, from Cornell's graduate SE course.
Overview
CS 6158 is a 3-credit Cornell graduate course taught by Assistant Professor Saikat Dutta. The Fall 2026 edition was renamed from Software Engineering in the Era of Machine Learning to Software Engineering in the Era of AI and adds material on how LLMs are changing software development. The schedule runs in two blocks. SE Basics (program analysis, software testing, security) comes first, followed by LLM Basics: intro to LLMs, fine-tuning and quantization, LLM agents, optimizations, evaluation of LLMs, speculative decoding and alignment, security of agents, and LLM/agent adaptation. The second half pairs classic SE problems with recent papers: dynamic analysis, flaky tests, formal verification, code translation, security, specification inference, fuzzing, debugging DL training, GPU kernel testing and verification, and testing DL libraries and compilers. Readings are linked from the schedule and include GEPA (ICLR 2026 oral), ProgramBench, SWE-smith, SWE-EVO, SWE-Gym, SPECBENCH on reward hacking in coding agents, information-flow control for agents, AutoVerus, CyberGym and Specula. Enrolled students present papers, post questions on Perusall and complete a semester research project in pairs plus 3-4 individual assignments. For readers outside Cornell, the free part is the dated schedule and paper list. Lecture slides go to Canvas and are not posted publicly, so this works best as a curated 2026 reading list on coding agents and AI-assisted verification.
At a Glance
- Topic
- Agentic
- Level
- Advanced
- Format
- Course
- Cost
- Free
- Duration
- 16 weeks (Aug 24 to Dec 7, 2026), two 75-minute paper sessions per week; the public reading list is self-paced
- Provider
- Cornell University (Saikat Dutta)
- Hands-on
- No
- Certificate
- None
What You’ll Learn
- ✓How LLM agents are evaluated on long-horizon software tasks with SWE-smith, SWE-EVO and ProgramBench
- ✓How reflective prompt evolution (GEPA) and test-time compute scaling compare with reinforcement learning for optimization
- ✓How information-flow control and probabilistic verification are used to secure AI agents
- ✓How LLMs generate proofs and formal specifications, through AutoVerus, Cobblestone, SpecGen and Specula
- ✓How LLM-driven pipelines translate whole repositories, including C to Rust and COBOL translation
- ✓How agents are applied to fuzzing and security vulnerability exploitation benchmarks such as CyberGym and ExploitGym
- ✓How SE techniques catch silent errors in DL training, GPU kernels and deep learning compilers
- ✓How SWE-Gym and SPECBENCH approach training coding agents and measuring reward hacking
Highlights
- •Reading list is current to mid-2026, with several arXiv papers from April to July 2026 alongside established ones
- •Covers both directions: AI for software engineering and software engineering for ML systems, including GPU kernels and DL compilers
- •Each session pairs a classic SE topic (flaky tests, fuzzing, verification) with an LLM-era paper on the same problem
- •Taught by a researcher whose own papers on flaky tests and ML testing appear in the readings
Who It’s For
Best For
- ✓Engineers building or evaluating coding agents who want the 2026 research landscape
- ✓Graduate students looking for SE and AI research project directions
- ✓Security and verification engineers exploring LLM-assisted proof and fuzzing work
Prerequisites
- •Software engineering or programming languages background (Cornell lists CS 3110, CS 4120 or CS 5154)
- •Machine learning or NLP coursework (Cornell lists CS 3780, CS 4740 or CS 4782)
- •Comfort reading research papers; strong Python and Java for the project work
FAQ
What is CS 6158: Software Engineering in the Era of AI (Cornell, Fall 2026)?
Cornell's graduate seminar CS 6158, taught by Saikat Dutta, on where software engineering and AI meet: using LLMs and agents for testing, verification, fuzzing and code translation, and applying SE techniques to ML systems. The public schedule and paper list suit engineers and researchers building or evaluating coding agents who want a vetted 2026 paper trail.
Is CS 6158: Software Engineering in the Era of AI (Cornell, Fall 2026) free?
CS 6158: Software Engineering in the Era of AI (Cornell, Fall 2026) is free to access.
What level is CS 6158: Software Engineering in the Era of AI (Cornell, Fall 2026) for?
CS 6158: Software Engineering in the Era of AI (Cornell, Fall 2026) is aimed at a advanced audience. Recommended background: Software engineering or programming languages background (Cornell lists CS 3110, CS 4120 or CS 5154), Machine learning or NLP coursework (Cornell lists CS 3780, CS 4740 or CS 4782), Comfort reading research papers; strong Python and Java for the project work.
How long does CS 6158: Software Engineering in the Era of AI (Cornell, Fall 2026) take?
Expect roughly 16 weeks (Aug 24 to Dec 7, 2026), two 75-minute paper sessions per week; the public reading list is self-paced. Most learners work through it at their own pace.
What will I learn from CS 6158: Software Engineering in the Era of AI (Cornell, Fall 2026)?
You'll learn: How LLM agents are evaluated on long-horizon software tasks with SWE-smith, SWE-EVO and ProgramBench; How reflective prompt evolution (GEPA) and test-time compute scaling compare with reinforcement learning for optimization; How information-flow control and probabilistic verification are used to secure AI agents; How LLMs generate proofs and formal specifications, through AutoVerus, Cobblestone, SpecGen and Specula; How LLM-driven pipelines translate whole repositories, including C to Rust and COBOL translation; How agents are applied to fuzzing and security vulnerability exploitation benchmarks such as CyberGym and ExploitGym; How SE techniques catch silent errors in DL training, GPU kernels and deep learning compilers; How SWE-Gym and SPECBENCH approach training coding agents and measuring reward hacking.
Topics
Sources
This page was written from 5 sources, 3 on domains other than cs.cornell.edu.