2026-08-18: KDD 2026 in Jeju, South Korea

From August 9-13, I attended the 32nd ACM SIGKDD Conference on Knowledge Discovery and Data Mining (KDD) in Jeju, South Korea, with my co-first author, Bang Nguyen. It was my first time traveling to East Asia, and the trip was full of new experiences and moments that I will cherish forever. 

The main reason for traveling to Korea was to present our paper, "ReplicatorBench: Benchmarking LLM Agents for Replicability in Social and Behavioral Sciences", which was accepted to the AI4Sciences track of KDD. We also presented the work at the SciSoc Agents and LLMs workshop. 

This work was a collaboration between the University of Notre Dame, Pennsylvania State University, Old Dominion University, and the Center for Open Science. 

Day 1

The first day started with the SciSoc Agents and LLMs workshop, where we presented our work.

It was a nice way to start the conference and also our first opportunity to present ReplicatorBench at KDD. The workshop was also a good introduction to the broader conference. A lot of the talks were closely related to LLM agents, social science, and scientific discovery, so it was interesting to see what other researchers are currently working on in these areas. 

In the afternoon, I also attended the lecture-style tutorial, When Models Leave the Training Distribution: A Tutorial on Out-of-Distribution (OOD) Detection. I was interested in whether ideas from OOD research could help us think about scientific hypothesis generation, particularly how we can encourage LLMs to generate novel ideas. The tutorial focused mostly on OOD detection in classification rather than text generation with LLMs, so it did not directly answer my questions, but it sent me down a path worth exploring. 

Day 2

On the second day, we had our poster session for the main conference. 

I also spent around seven hours volunteering as part of receiving a student travel grant from the conference. It made for a long day, but the travel support was a big help in making the trip to Korea possible. 

One of the best parts of the conference was also getting to meet a lot of new people. I met PhD students from different universities who are working on similar problems, as well as recruiters and researchers from industry. These conversations were useful from both a research and career perspective. Conferences like KDD bring together people from academia and industry in the same place, so it was an excellent opportunity to learn about what others are working on, exchange ideas, and make new connections.

Day 3

The third day started with Jeff Dean's keynote, which was one of the talks I was most excited about. He was the co-founder of Google Brain, served as Chief Scientist at Google DeepMind, and was technical co-lead for Gemini. He has had a career that I have looked up to for a long time. He walked the audience through his career and many of the systems he worked on that helped lay the foundation for modern AI. Hearing this story directly from him was the highlight of the conference. 

He also shared the emotional hourly schedule of his final day at Google before leaving after almost three decades. What caught my attention the most was his new startup, Discovery Loop, which focuses on using agentic AI to accelerate scientific discovery and hypothesis generation. This is strongly connected to my current research interests, so it was exciting to see someone with his background moving in this direction. 

I was also hoping to ask him about scientific hypothesis generation and his thoughts on generating novel hypotheses beyond the training distribution, but as you can tell from the picture on the left, I wasn't the only one hoping to talk to him. 

Day 4

On the fourth day of KDD, we had the opportunity to listen to Jingren Zhou's keynote. He is currently Senior Vice President and Chief AI Architect at Alibaba. 

His talk focused on the role of LLMs across the data and training pipeline, covering topics from data mining and processing to agentic systems, including Qwen models being developed at Alibaba. 

It was particularly interesting to hear how these ideas led to the development of Qwen, currently the most downloaded open-source LLM family, and how much of the current progress in AI is moving beyond just the model itself toward larger systems involving tools, data, memory, and agents.

Looking back

Overall, KDD 2026 was a great experience. 

The main goal of this trip was to present our work, but the conference ended up being much more than that. I got to visit a completely new part of the world, listen to researchers who have had a major impact on the field, and meet many new people from both academia and industry. 

For me, the networking part was especially valuable. As a PhD student, it is easy to spend most of your time focused on your own research problems, but being at a conference like KDD gives you a much better sense of what the wider research community is working on and where the field is heading. 

I would also like to thank my advisor, Dr. Jian Wu, for making this trip possible. 

It was a busy few days, but definitely a trip I will remember. 

~Dominik Soós