TeaP2025 ended

Helping organizing a conference

TeaP2025 (2025-03-09 to 2025-03-12, Frankfurt am Main, Germany) ended successfully. Here are some thoughts to share:

LLM-based conference organization

The idea was: “Can we use the large-language model [LLM] embedding to measure semantic distance between conference abstracts?” 🤔

Large-language model?

You know, things like ChatGPT, Claude, Gemini 🤖. The things that parse strings of text tokens and generate fake text that sounds plausible. Maybe something similar to our “language center” in our brains does when we are extremely drunken. And also things that generate your fake homework essays and correct typos of this post.

Embedding?

It’s what non-philosophers think a “representation” would be like for an artificial neural network like a LLM. This is defined for a word, a sentenece, a paragraph, or a whole writing given as a set of 4000+ numbers (i.e., a vector) for an OpenAI model (text-embedding-3-large). See also: https://platform.openai.com/docs/api-reference/embeddings/create

Clustering and organizing

Once you define the distance between abstracts, then you can perform clustering on that.

We wanted to organize talk sessions while satisfying the following criteria:

  • Find near-optimal clustering with a maximaum of six abstracts per session.

  • Once clusters are found, maximize distances between parallel sessions.

We wanted to organize poster sessions while considering these conditions:

  • Find near-optimal clustering while maintaining the number of clusters manageable for manual browsing on the conference web page (k < 10).

  • Ensure that poster presenters to view other posters on similar topics. To this end, the closest three posters are assigned to different days.

  • Optimize attendees’ walking paths by arranging posters so that their semantic similarity aligns with their physical proximity.

Want to know more? [PDF]