Organising a seminar on Large Language Models

Last week I wrapped up a seminar series on LLMs that I organised at Aalto University and I thought of sharing in this article some of the things I learned.

Incipit

As a background, the seminar ran from January to June 2025, with weekly presentations of recent LLM papers given to and for Aalto staff members that were interested in the topic (mostly PhDs and postdocs). Typically a session would last 1 hour, with around 20 minutes allocated to the presentation and the rest of the time for discussions (which were encouraged throughout the presentation itself). The participants took turns in presenting and I kept things running. As organiser, my main duties were: 1) inviting people to the seminar and kick things off, 2) find presenters among participants, 3) booking the room for the event, 4) advertising the seminar on Slack, and in the department via a newsletter and ads on the department screens and 5) keeping the website up-to-date. Finally, I facilitated the in-person discussions.

This was the second iteration of the seminar (I organised one in 2024 with the help of Minttu Alakuijala), and the main reason why I organised it was to bring together researchers at my university who were either doing active research or being interested in LLMs. I had multiple motivations behind this, such as learning from papers and other people’s thoughts on them, networking, socialising and having deep discussions that would hopefully turn me into a better researcher.

The story

As organiser, I had to face some questions before launching the seminar: Who would decide the topics to treat? How much commitment would I require from the participants? Who was the target audience and what was the minimum level of expertise required to participate? And finally, how to attract and retain enough people to keep the seminar dynamic and engaging?

Being at the second edition, I knew some things: First, I wasn’t there only to do a service to others, but I cared about getting value out of it too. This motivated me to treat only cutting-edge topics and have them overlapping significantly with my own interests; thus I aimed at reaching specifically PhD students and above. Second, I knew academics at all levels are super busy and overwhelmed with work, so that made me opt for a low commitment seminar, where people could attend even just one session and didn’t need to necessarily promise to give a presentation in order to attend. My hope was to extend the reach of the seminar to virtually everyone who could be interested in the topic in the department, and then being able to retain the most passionate people through the quality of the seminar itself. Finally, I opted for letting everyone pick its own paper to present, limiting myself to some suggestion if they didn’t have anything in mind. This way I thought it was probably further reducing the burden on the participants, as they could present some paper they had to read for their own research and not start from scratch on the presentation.

The launch itself was quite tricky: I didn’t know who would show up, nor how to gather feedback about who would like to attend. I had a Slack channel with people already in it from the year past, then I spread the invitation to Master’s students who took a seminar course on LLMs given by my supervisor and I the previous semester. I also reached out extensively to the Electrical Engineering department, where I heard that 3 research groups could have some interest in LLMs. Finally, I put up a website to make the seminar more official, and advertised it on the department newsletter and screens of the coffee rooms. When everything was starting to be in place, I decided to launch and present a paper that caught my attention, Byte Latent Transformer from Meta. In retrospect, not the best idea, as it was quite complicated and not as consequential as I thought before reading it, but such is life. More than 10 people showed up, among which 2 professors (to my surprise) and we had a good presentation and discussion and I felt encouraged from that start. Also, I managed to find a presenter for the week after and started contacting some of the participants to ask if and when they could present and draft a schedule for the following sessions.

However, the week after, there were maybe 3 or 4 people attending, and I felt quite disappointed by the drop. Furthermore, I had no presenter for the week after that, and the whole thing felt overwhelming. Luckily, that week DeepSeek-R1 came out and a friend of a friend volunteered to put up a presentation, so just 10 days after the publication of the paper we were having a seminar on that. The session went great, and something like 13 to 15 people were attending. This gave me some leeway to find presenters for the entire month ahead and things got more stable. However, the attendance kept fluctuating, and many people came once (even to present) and then never came back, which felt like a real bummer. Also the amount of people joining the Slack channel kept on growing (currently 93 members), but that didn’t really translate to new participants.

I found it really helpful to reach out to most participants to get some feedback and get a better understanding of why they weren’t coming at certain times (without asking it straight up, but more by just asking how their research was going and stuff like that). Many times it would be deadlines or holidays, whereas other times I just had no idea. I was also contacted twice from people who wanted to present something dear to them and just heard about the existence of the seminar one way or another, which I have to say was quite cool (both turned out to be great presentations).

Overall, I was never 100% satisfied by the sessions, as my expectation bar was quite high, in terms of presentations, discussions, participation and also follow-up engagement. Between May and June there were few interruptions, between public holidays, the NeurIPS deadline, and some time off for vacation; the interruptions made me feel less invested in the seminar and I saw things stagnating a bit. The final signal was the announcement of big renovations to the department for two months and a half, with the encouragement to all the staff to work from home. After a couple of weeks I decided that it was a good time to stop the seminar and start investing my time in other activities, such as this blog and searching for what to do after the PhD.

Lessons learned

Out of this seminar I grew not only as a researcher, but also as a person. Looking back, I see that I changed my mind on certain matters, such as why it’s worth running a seminar, what are the biggest challenges, and what is possible to do on your own in academia beyond your own duty.

The biggest value for me was in the experience of organising the seminar itself, the confidence it built in me to take initiative and turn my ideas into action, and all the skills I practiced in the process. Instead, from the seminar proper, I would say that it was a great opportunity to network with 20-30 people across two departments at my university, which gave me a good intuition about the character of each of them and might inform future collaborations of mine (I’m not planning any for the time being, as I’m about to graduate from the PhD, but you never know about the future!) Then, the topic chosen from the participants, ranging from explainability, unlearning and safety, to agent protocols and AI scientists, also brought some fresh perspectives for me, beyond the content of the papers themselves. Finally, some discussions also influenced my thinking, in minor or major ways; for example I got into deep discussions about AI safety with one participant and started considering it more seriously, and I also got to hear more insights about DeepSeek-R1 and the Chinese ecosystem from the Chinese student who presented the paper.

Some of the biggest challenges I encountered were the time investment it required and the emotional toll it took on me at times. In terms of time, the bulk of it went into the initial organization: drafting a lift of people who could be interested, creating the website, the recurrent zoom link, understanding how to create a new event for each seminar session, sending emails to the research groups who could possibly be interested, preparing the ads for my department and then preparing my own paper presentation; however, this was expected. What was not expected was the time I spent in steadying the course of the seminar when I couldn’t find a presenter or I saw the affluence decreasing. In those circumstances I mostly did two things: spent time thinking and revising the whole seminar pipeline, to try and intervene in the most promising parts, and reaching out to a lot of participants, to know how they were doing, whether they had any feedback about the seminar, if they wished to present in the future, which topics or formats were they appreciating the most and so on.

In terms of emotional toll, mostly it’s hard not to feel like your worth is a reflection of the success (or insuccess) you are having. For instance, it’s hard to be confident that you are organising something worth attending and it’s hard not to take it personally when people who manifest interest in what you are organising don’t show up. It’s a good exercise in resilience, self-compassion and open-mindness.

Finally, one thing that really surprised me is that a good chunk of people were more likely to come only the time they promised to present than to attend other sessions without ever presenting (it happened roughly 5 times). I am still trying to square this one out in my mind. I would have understood if it was to present their own research, as you have some sort of personal return that offsets the extra time investment, but other than that, maybe it suggests that for some people some form of commitment was needed to come at the seminar.

Overall, it’s a great formative experience on so many fronts and certainly one I am grateful of having had, yet maybe not the best return on investment if all you care about is reading papers you like and having interesting discussions about them.