Master Seminar - Networked AI Systems (NAIS) - Chair of Connected Mobility

Master Seminar - Networked AI Systems (NAIS) (Winter 2026/27)

This seminar explores Networked Systems for AI workloads from both directions: systems and networks for AI (e.g., GPU cluster networking, LLM training/serving infrastructure) and AI for systems and networks (e.g., learned network optimization and autonomous management).

This Master’s seminar explores these systems from the perspectives of underlying network infrastructure, system design, and their interaction with services like deep learning models. These systems include: LLMs training supported by GPU clusters, LLM inference caching, LLM infrastructure design, Parallel and Collective communication. We read and discuss recent papers from top venues such as SIGCOMM, NSDI, MLSys, OSDI, and SOSP.

Unlike a classic review-based seminar, this seminar is discussion- and presentation-based. There are no written paper reviews. Instead, every participant reads the assigned paper before each session, and we discuss it together in a round-table format: the presenter introduces the paper, then the group dissects its ideas, assumptions, weaknesses, and — most importantly — what could come next. The goal is not only to understand state-of-the-art research, but to generate new research ideas upon it.

Ideas that emerge from these discussions can directly grow into Guided Research, Master’s thesis, or internship topics at our chair. Strong participants with promising ideas will be actively considered for such positions.

The topics may include but are not limited to:

Networking for large-scale ML training (e.g., collective communication, congestion control for AI workloads, GPU cluster fabrics, RDMA)
LLM serving and inference systems (e.g., scheduling, caching, disaggregation, latency/throughput trade-offs)
ML for networking or Systems (e.g., learned congestion control, traffic engineering, network telemetry and diagnosis, application layer routers)
AI at the edge and in the device–edge–cloud continuum
Resource management and orchestration for AI workloads (e.g., GPU sharing, job scheduling, energy efficiency)
Emerging interconnects and hardware-network co-design for AI systems
Sustainability and energy efficiency of AI systems and its networking implications

Registering for this seminar follows the usual process of the Matching System of the Department of Computer Science

Registration & Motivation Email

Places in this seminar are very limited in order to ensure a high-quality discussion environment. Prospective participants can increase their chance of being accepted by sending a short motivation email to show your potential contribution to the seminar:

Subject: “[Seminar NAIS 26/27] Motivation - Your Name”
Content: max 200 words on why you want to take this course (e.g., your background, your interest in networked AI systems, or what you hope to get out of the discussions). For example: " I want derive a research idea for my Master’s thesis from this seminar, and I am particularly interested in the intersection of ML and networking. I have experience with distributed systems and have implemented a custom congestion control algorithm in a previous project."

Moodle page

Master Seminar - Networked AI Systems (NAIS) [IN2107, INxxxx]

Time and location

The seminar sessions will be held on-site in.
You can also subscribe the calendar here Seminar - NAIS or iCal format
There will be no pre-course meeting but you may refer to our introduction slides here.
Attendance at all sessions is mandatory — discussion is the core of this seminar. Only well-excused absences (e.g., illness with certificate) are accepted and may lead to an online participation.

Time: 16:00-18:00

[Oct 14]: Introduction & QA Session — basic information, FAQs, and paper assignment. Will be short. MI 01.07.023
[Oct 21]: Seminar Session #1 Room: MI 01.07.023
[Nov 04]: Seminar Session #2 Room: MI 00.13.036 (noting the room change❗)
[Nov 18]: Seminar Session #3 Room: MI 01.07.023 Plus course survey time (pls do do the course survey which helps this seminar grow better)
[Nov 29]: Write-up DDL（No Session)
[Dec 02]: Grading (No Session)

Objective

Upon completion of the seminar, the students will

have broadened their knowledge of the current state of Networked AI Systems research and its open gaps
gain critical thinking and analytical skills through in-depth, round-table discussion of research papers
learn to articulate, defend, and refine technical arguments in an open scientific discussion — a core skill for research
develop and sharpen their own research ideas, potentially seeding a Guided Research, Master’s thesis, or internship topic

Languages of Instruction

English

Teaching and Learning Method

Workload

Each participant will be required to:

Read all papers of each discussion session (max 4 papers per session) before the session (about 1–3 hours per session) — careful reading is the foundation of good discussion
Present 1 paper during the semester (about 2-5 hours workload, 40% of grade)
Actively participate in the round-table discussions of all sessions (60% of grade) — quality of contributions counts, not just quantity
Submit ONLY one 1-page write-up before the deadline.

The write-up

Once during the semester (deadline: one week after the last session), each participant submits a single 1-page write-up using this LaTex Template. It is not a paper review or summary. Pick one focused idea inspired by the seminar — a critique you developed during discussion, a new research direction, an extension of a paper, or a question worth investigating — and present it in a formal, well-argued way. Don’t just talk cheap; back up your claims with evidence. Choosing what not to write is part of the exercise. Strong write-ups are a natural starting point for thesis or Guided Research proposals.

This is not a formality-driven assignment submission, nor is it a credits game. If you are only looking to take a seminar and earn credits, I’m sorry, but this is not the right place for you. We very much hope to receive some insightful thoughts and forward-looking perspectives. So please feel free to express your individuality and write something different, rather than offering dull clichés. Of course, all of our instructors will read your submission carefully, and it may even lead to some deeper exchanges.

Workflow

Before Course Starts: Participants state their paper preferences during the introduction session and the lecturers assign each paper to a presenter and to a specific session.
Before Formal Sessions: Everyone reads the session’s papers in advance — there are no passive listeners in this seminar.
During Formal Sessions: In each session (max 4 papers), the presenter introduces the paper (~15 min), then the group discusses it round-table style (~15 min per paper): strengths, weaknesses, hidden assumptions, follow-up ideas, or future directions. Everyone can raise their questions. Debate is encouraged :)

(The exact numbers may slightly vary due to the number of registered students.)

What counts as “active engagement”?

Your discussion grade reflects the quality of your contributions, not how often you speak. Intuitively, contributions like these count positively:

Prepared questions: “The paper claims X, but their evaluation only covers Y — would this hold under Z?”
Constructive critique: pointing out a hidden assumption, a missing baseline, or a threat to validity — and suggesting how it could be addressed
Connections: linking the paper to another paper, technique, or session (“this is essentially the same trade-off as in last week’s paper, but at a different layer”)
New ideas: proposing a follow-up experiment, an extension, or an application of the idea in a different setting
Moving the discussion forward: building on others’ points, respectfully disagreeing with arguments (not persons), summarizing where the debate stands

Merely attending, or repeating the presenter’s summary, does not count as engagement. We will be genrous in grading as long as you are generous in contributing to the discussion.

Why take this seminar?

Engage deeply with state-of-the-art research from SIGCOMM, NSDI, and other top venues — without the overhead of formal review writing
Practice the discussion culture of real research groups and program committees, the minimum instructor-to-student ratio is 1:4.
Develop your own research ideas with direct feedback from peers and instructors; all your peers are strongly motivated.
Potential computing infrastructures and workspace in the chair for your hands-on (if you have) and close supervision from the instructors.
Outstanding participants may be offered Guided Research, Master’s thesis, or internship topics growing out of the seminar discussions.

AI Policy

You may use LLMs to learn technical details, learn background knowledge, and improve your understanding of the papers.
Your contributions in the discussion sessions are inherently your own — come prepared to think on your feet.
The final 1-page write-up must be written entirely in your own words. AI-detection tools will be used to screen submissions; AI-generated content will be regarded as cheating and result in a failing grade.
AI coders are encouraged for any practice to understand the papers or algorithms, or for any code backing up your final write-up.

Contact

Please subject your email with “[Seminar NAIS] - your query”, to ALL of us, thanks. We can have brief conversations or discussions in emails to get to know each other better.

You can also give feedbacks or interact with us or (perhaps) previous students at this notion page.

📧Click to Send an Email to instructors

Wei Geng wei.geng@tum.de — Office Hours: 13–13:30 Wed (Email preferred)
David Guzman david.guzman[at]cit.tum.de
Hyerin Kim hyerin.kim[at]tum.de

To top