Nicola Branchini

"Accept (spotlight)" at ICML'26 😎 Our paper brings particle filters back to life: autoregressive diffusion models + posterior sampling yield optimal proposals for Bayesian filtering, scaling up to GenCast-sized systems. arxiv.org/abs/2605.20028 w/ Thomas Savary and @francois-r…

[2605.20028] Training-Free Bayesian Filtering with Generative Emulators arxiv.org View on Bluesky →

With friends at the University of Warwick (in particular, Rocco Caprio and @adriencorenflos.bsky.social), we've recently arXived some work (arxiv.org/abs/2605.30253) on a method for approximate inference known as "Coordinate Ascent Variational Inference", or "CAVI" for short. …

Wasserstein Contraction of Coordinate Ascent Variational Inference arxiv.org

AI Weekly's analysis →

• The paper establishes Wasserstein contraction of coordinate ascent variational inference without assuming global strong log-concavity of the target.
• The conditions are a functional smoothness of the optimality maps plus a transportation-information inequality at their fixed points.
• Covered models include Ising and Curie-Weiss, Bayesian Gaussian mixtures, high-dimensional Bayesian probit regression, and Pólya-Gamma logistic regression.

Read full analysis →

View on Bluesky →

Mehdi defends our recent preprint on rho-posteriors and their variational approximations at ISBA @Nagoya Link to the preprint: arxiv.org/abs/2601.07325

Robust Bayesian Inference via Variational Approximations of Generalized Rho-Posteriors arxiv.org

AI Weekly's analysis →

• Khribch and Alquier introduce a rho-tilde-posterior that swaps the supremum over competitor parameters for a softmax aggregation.
• The construction yields PAC-Bayesian finite-sample oracle inequalities with explicit convergence rates that survive model misspecification and data contamination.
• Those guarantees extend to variational approximations, with computational cost the authors report as comparable to standard variational Bayes.

Read full analysis →

View on Bluesky →

🧵 Preprint alert! Introducing Augmented Gaussian sum filters (AGSF), a novel class of Bayesian filtering algorithms which unifies Gaussian sum (GSF) and particle filters (PF) by interpolating continuously between them, while being robust to common failure modes. arxiv.org/abs/…

A Gaussian Sum Filter for Unifying Gaussian and Particle Filters arxiv.org View on Bluesky →