Publications

S. Bessa, P.-E. Mazare, G. Szilvasy, H. Touvron, D. Picard, M. Douze, H. Jégou

e-means: Outperforming k-means (in One-Tenth the Epochs)

ICLR 2027 (submission)

e-means is a modern optimization rule for k-means, built around mini-batches with an exponential moving average rule.

To be released publicly in early November 2026.

S. Bessa, P. Fernandez, G. Szilvasy, M. Douze, H. Jégou

Cluster with Auctions for Vector Search

NeurIPS 2026

S. Bessa, D. Dabert, M. Bourgeat, Q. Cappart, L. Rousseau

Learning Valid Dual Bounds in Constraint Programming: Boosted Lagrangian Decomposition with Self-Supervised Learning

AAAI 2025