Category Theory
Zulip Server
Archive

You're reading the public-facing archive of the Category Theory Zulip server.
To join the server you need an invite. Anybody can get an invite by contacting Matteo Capucci at name dot surname at gmail dot com.
For all things related to this archive refer to the same person.

Stream: community: our work

Topic: Peva Blanchard

Peva Blanchard (Nov 13 2025 at 08:58):

Our paper Generalizing while preserving monotonicity has been accepted at NeurIPS 2025. It's been a long time since I last published a paper, so I'm quite happy. And this zulip is partly responsible for getting me back in the mood of doing research again.

In short, we study models (variants of Bradley-Terry models) that learn preferences from comparisons between alternatives. We want these models to be monotone: if you say you prefer $a$ over $b$, the score of $a$ should increase and the score of $b$ should decrease. Surprisingly, this property is not always satisfied. In our work, we assume that the alternatives are represented by vector embeddings, yielding an embedding matrix $x$, and that the score is a linear function of embeddings. We found a sufficient condition (diffusion embedding) on $x$ for monotonicity to hold, namely that the inverse of the Gram matrix $x^Tx$ is super-laplacian. Intuitively, with a diffusion embedding, score propagates like heat: stating that you prefer $a$ over $b$ is like installing a heat pump that heats $a$ at the expense of $b$.

So it is not directly related to CT. But in future work, I would like to understand a bit more these diffusion embeddings, probably using ideas from open graphs.

Hans Riess (Nov 21 2025 at 10:49):

Peva Blanchard said:

Our paper Generalizing while preserving monotonicity has been accepted at NeurIPS 2025. It's been a long time since I last published a paper, so I'm quite happy. And this zulip is partly responsible for getting me back in the mood of doing research again.

I haven't had the time to fully read the paper yet, but this looks very interesting. We wrote a paper on preference dynamics of networked agents. I interact more with the robotics and control community, but I think the framework we established could be useful for RLHF for networks of LLM agents.

Peva Blanchard (Nov 27 2025 at 19:18):

Interesting I'll have a look at your paper.