CS 394E: Contemporary Topics in LLMs (KAUST, Fall 2026)
by KAUST (Marco Canini)
A KAUST graduate course on LLM systems: agents, hallucination, evaluation, RAG, memory, MCP/A2A and serving with PagedAttention.
Overview
CS 394E Contemporary Topics in LLMs is a fall 2026 graduate course in KAUST's computer science department, taught by Marco Canini with teaching assistants Norah Alballa and Kaihua Liang. It is organized into four modules over fifteen weekly three-hour sessions. LLM and Agents Foundations covers n-gram models, tokenization and embeddings, attention and transformers, then post-training, alignment and tool use with ReAct and s1 test-time scaling. Reliability and Evaluation covers hallucination detection with semantic entropy, why models hallucinate, benchmarks, LLM-as-a-judge, Chatbot Arena and The Leaderboard Illusion. Scalable LLM Systems covers RAG and AutoGen-style multi-agent frameworks, why multi-agent systems fail, MemGPT, Recursive Language Models, MCP and A2A, and serving with Orca, PagedAttention, Mooncake and RouteLLM. Domain Applications covers LLMs for scientific discovery, healthcare and education, world models, and vision-language-action agents such as OpenVLA and pi-zero. The primary text is selected chapters of Jurafsky and Martin's Speech and Language Processing (3rd edition draft, free online). Students write a summary of at most 500 words for each assigned paper, complete three one-week builds (fine-tuning, an agent loop, RAG) and run a team research project. Slides are posted after each class, and the reading list was last updated 6 October 2026.
At a Glance
- Topic
- Agentic
- Level
- Advanced
- Format
- Course
- Cost
- Free
- Duration
- 15 weeks (31 Aug to 7 Dec 2026), one 3-hour session per week; slides posted after each class (weeks 1-6 up as of 11 Oct 2026)
- Provider
- KAUST (Marco Canini)
- Hands-on
- Yes — code/exercises
- Certificate
- None
What You’ll Learn
- ✓Detect hallucinations with semantic entropy and explain why language models hallucinate
- ✓Evaluate models with benchmarks, LLM-as-a-judge and Chatbot Arena, and spot leaderboard bias
- ✓Diagnose why multi-agent LLM systems fail using published failure taxonomies
- ✓Design agent memory with MemGPT ideas and connect tools through MCP and A2A
- ✓Explain LLM serving with Orca continuous batching, PagedAttention and KV-cache-centric designs
- ✓Apply test-time scaling and ReAct-style tool use after post-training and alignment
- ✓Read vision-language-action papers such as OpenVLA and pi-zero for embodied agents
Highlights
- •Covers serving systems (Orca, PagedAttention, Mooncake, RouteLLM) alongside agents, which few LLM courses pair
- •Readings run through 2026, including Recursive Language Models and The Leaderboard Illusion
- •Two-track paper assignments let you pick a theory or an applied reading in several weeks
- •Free primary text: Jurafsky and Martin's SLP3 draft, updated August 2026
Who It’s For
Best For
- ✓AI engineers building multi-agent or RAG systems who want the research behind them
- ✓Inference and platform engineers who want serving papers in an LLM context
- ✓Graduate students choosing an LLM systems research project
Prerequisites
- •Comfortable Python programming and basic deep learning
- •Ability to read and summarize research papers
- •Distributed systems or NLP tooling experience helps but is not required
FAQ
What is CS 394E: Contemporary Topics in LLMs (KAUST, Fall 2026)?
A KAUST graduate seminar for engineers who already know basic deep learning and want a paper-driven tour of how LLM systems are built and evaluated. It runs from n-grams and transformers through hallucination detection, LLM-as-a-judge, RAG, multi-agent failures, agent memory, MCP and A2A, and LLM serving.
Is CS 394E: Contemporary Topics in LLMs (KAUST, Fall 2026) free?
CS 394E: Contemporary Topics in LLMs (KAUST, Fall 2026) is free to access.
What level is CS 394E: Contemporary Topics in LLMs (KAUST, Fall 2026) for?
CS 394E: Contemporary Topics in LLMs (KAUST, Fall 2026) is aimed at a advanced audience. Recommended background: Comfortable Python programming and basic deep learning, Ability to read and summarize research papers, Distributed systems or NLP tooling experience helps but is not required.
How long does CS 394E: Contemporary Topics in LLMs (KAUST, Fall 2026) take?
Expect roughly 15 weeks (31 Aug to 7 Dec 2026), one 3-hour session per week; slides posted after each class (weeks 1-6 up as of 11 Oct 2026). Most learners work through it at their own pace.
What will I learn from CS 394E: Contemporary Topics in LLMs (KAUST, Fall 2026)?
You'll learn: Detect hallucinations with semantic entropy and explain why language models hallucinate; Evaluate models with benchmarks, LLM-as-a-judge and Chatbot Arena, and spot leaderboard bias; Diagnose why multi-agent LLM systems fail using published failure taxonomies; Design agent memory with MemGPT ideas and connect tools through MCP and A2A; Explain LLM serving with Orca continuous batching, PagedAttention and KV-cache-centric designs; Apply test-time scaling and ReAct-style tool use after post-training and alignment; Read vision-language-action papers such as OpenVLA and pi-zero for embodied agents.
Topics
Sources
This page was written from 3 sources, 1 on domains other than sands.kaust.edu.sa.