What you will be able to do
Trace a decoder-only Transformer, design reproducible evaluations, build trustworthy RAG and agent systems, and reason about training and serving trade-offs.
2026 schedule publishedLast updated 25 August 2026
M2 DISS · Fall 2026 · 6 ECTS
MLTAFoundation Models and Agentic Systems
Modern machine learning through foundation models—following the full path from generation and evaluation to retrieval, agents, systems, and multimodality.
Instructor
Chao Zhang
Junior Professor Chair in Computer Science
Université Claude Bernard Lyon 1
33focused sessions
36hlectures
24hlabs & project
50/50exam / practical
01 · Course overview
MLTA is a graduate-level course on modern machine learning with large language models as its principal technical thread.
We move through the complete lifecycle of an LLM system: how text becomes tokens and logits, how models are trained and aligned, how external knowledge is retrieved, how agents plan and act, and how the resulting systems are evaluated and served.
Mechanisms, empirical evidence, controlled comparison, failure analysis, and system trade-offs are central throughout. Products and frameworks appear as case studies rather than ends in themselves.
What this course is not
A general introduction to classical machine learning. Regression, trees, clustering, calculus, probability, and linear algebra are prerequisites or reviewed just in time.
02 · Syllabus
- 01
Sessions 01–05
Foundations, Prompting & Evaluation
How do models generate—and how can we compare them reliably?
- 02
Sessions 06–08
Transformer Architecture
How does a modern decoder represent and transform a sequence?
- 03
Sessions 09–12
Pretraining, Alignment & Model Updates
How are capabilities learned, aligned, and revised?
- 04
Sessions 13–16
Retrieval & RAG
How can external knowledge make answers verifiable?
- 05
Sessions 17–24
Reasoning, Agents & Self-Improvement
How do models plan, act, recover, and learn safely?
- 06
Sessions 25–29
Efficient Training & Inference
How do we make model workloads efficient and dependable?
- 07
Sessions 30–33
Foundation Models Beyond Text
How does the paradigm extend to structured and multimodal data?
03 · Calendar
Lecture · 09:45–13:00 Practical / project · afternoon Examination
Part 1Foundations, Prompting & EvaluationPart 2Transformer ArchitecturePart 3Pretraining, Alignment & Model UpdatesPart 4Retrieval & RAGPart 5Reasoning, Agents & Self-ImprovementPart 6Efficient Training & InferencePart 7Foundation Models Beyond Text
- Wednesday · Week 1
- 01Generative AI Foundations
- 02Tokenization, Embeddings, and Generation
- 03Prompt Engineering I: Instructions, Examples, Constraints
No afternoon block - Wednesday · Week 2
- 04Prompt Engineering II: Decomposition and Structure
- 05LLM Evaluation Foundations
- 06Self-Attention
No afternoon block - Wednesday · Week 3
- 07Transformer and Decoder-Only Language Models
- 08Modern LLM Architecture
- 09Pretraining, Data, and Scaling Laws
Lab / project block · activity mapping TBC
- Wednesday · Week 4
- 10Alignment: SFT and Reinforcement Learning
- 11Post-Training and Forgetting
- 12Model Editing and Model Merging
Lab / project block · activity mapping TBC
- Wednesday · Week 5
- 13Retrieval Foundations
- 14Vector Search and Indexing
- 15RAG Architectures
Lab / project block · activity mapping TBC
- Wednesday · Week 6
- 16Advanced RAG
- 17Reasoning, Planning, and Test-Time Scaling
- 18AI Agent Fundamentals
Lab / project block · activity mapping TBC
- Wednesday · Week 7
- 19Context and Harness Engineering
- 20Agent Memory
- 21Self-Correction and Verification
No afternoon block - Wednesday · Week 8
- 22Synthetic Experience and Self-Improving Agents
- 23Agentic RL
- 24Agent Evaluation, Reliability, and Security
Lab / project block · activity mapping TBC
- Wednesday · Week 9
- 25Training Systems and PEFT
- 26LLM Inference Mechanics
- 27Long-Context Inference
Lab / project block · activity mapping TBC
- Wednesday · Week 10
- 28Efficient Serving
- 29Serving Stack: vLLM, SGLang, and Caching
- 30Tabular Foundation Models
Lab / project block · activity mapping TBC
- Wednesday · Week 11
- 31Text-to-SQL and Data Agents
- 32Graph and Time-Series Foundation Models
- 33Annual Frontier
Lab / project block · activity mapping TBC
- Wednesday · Examination
Written examination
Coverage and format details to be confirmed.End of course
TBC
The dates, times, rooms, and morning session mapping are confirmed from the official timetable. The allocation of the twelve practical activities across the eight afternoon blocks remains to be confirmed.
04 · Practical work
- 01
TBC
- 02
TBC
- 03
TBC
- 04
TBC
- 05
TBC
- 06
TBC
- 07
TBC
- 08
TBC
- 09
TBC
- 10
TBC
- 11
TBC
- 12
TBC
05 · Assessment
50%
Written examination
Mechanisms, experiment and system design, trade-offs, failure diagnosis, and interpretation of evidence.
50%
Practical work
06 · Logistics
- Course
- Machine Learning Techniques and Applications: Foundation Models and Agentic Systems (MLTA)
- Programme
- M2 DISS · 6 ECTS
- Instructor
- Chao Zhang
- Department
- Département d'Informatique
Université Claude Bernard Lyon 1 - Primary venue
- Nautibus TD001 / TD005
Ground floor - Contact
- chao.zhang@univ-lyon1.fr
- Course materials
- Available on Moodle – Lyon 1
- Prerequisites
- Python, machine learning and deep learning foundations, probability, linear algebra, and calculus.