Benjamin Rozonoyer

Excited to share that our paper *Autoregressive Ranking: Bridging the Gap Between Dual and Cross Encoders*, is now available on arXiv! arxiv.org/pdf/2601.05588

Paper link 📎: openreview.net/pdf?id=HFS800r…

Excited to present our NeurIPS paper "Learning Representations for Hierarchies with Minimal Support" at the morning poster session on December 12! Stop by at poster #3500 in the East Building!

We see robust performance and fast convergence on synthetic hierarchies and the Medical Subject Headings taxonomy, with up to a ⚡ 99% reduction ⚡ in the size of the training example set.

🚀 We show that we can train node embedding models with greater efficiency using just this reduced set of examples, provided that our model has an appropriate inductive bias (in our case, transitivity bias).

💡 Applying this insight to hierarchies (transitively-closed directed acyclic graphs), we present an algorithm to compute the provably minimal set of edges to logically identify a graph among all hierarchies.

If we know that the graph has a certain structural property, then only a small subset of positive and negative edges might be required to uniquely identify that graph among all other graphs with that property.

When learning geometric representations of graphs, the model needs to see both “positive” edges from the graph, as well as “negative” edges from the edge complement.