Budapest, HU · PhD researcher · est. 2025–2029

Miklós Máté Badó

PhD researcher in the Department of Artificial Intelligence at Eötvös Loránd University and NIPG.

I'm interested in mean field games and, more broadly, in how mathematical structure — gradient geometry, curvature, continuous-time formulations — can be exploited to make principled inference, attribution, and uncertainty quantification tractable at the scale modern models actually live at. My most recent published work, Clustered Influence Functions (ICML 2026), is an amortized subset-influence oracle for high-query workloads. Before the PhD I worked on gesture-based human–machine interaction at NIPG.

News

Selected Publications

Research

My research sits at the intersection of optimization, applied probability, and machine learning. The questions I find motivating cluster around how mathematical structure — gradient geometry, curvature, and the way high-dimensional learning problems inherit structure from continuous formulations — can be used to make principled diagnostics, attribution, and uncertainty estimates tractable at the scale modern models actually run at.

My current PhD work is on mean field games — large-population stochastic systems whose macroscopic behaviour is described by coupled forward–backward partial differential equations. I'm interested in inference and uncertainty quantification under these dynamics, and in the algorithmic side of how continuous-time formulations expose structure that purely discrete or finite-population formulations can't.

My most recent published work is Clustered Influence Functions (CiF, ICML 2026, with Kristian Fenech). The classical influence function asks how a trained model would change if a subset of training data were removed — useful for data debugging, unlearning, mislabel detection, and large-$K$ cross-validation — but the per-query inverse Hessian-vector product makes it impractical for the high-query workloads it would be most useful in. CiF amortizes by clustering training gradients, solving the damped Gauss–Newton system once per cluster mean, and answering arbitrary subset queries by linear recombination. The method preserves exact damped-GGN semantics — only the operand space is compressed — so it stays complementary rather than competitive with projection-based attribution methods.

Before the PhD, my BSc work at NIPG centred on gesture-based human–machine interaction: how people and large language models can communicate solely through pragmatic gestures, and how to evaluate that empirically. That line produced the Keep Gesturing paper and a TDK first-place award.

Experience

Awards

Outside the lab

Trail running, climbing, scouting. Race writeups and the occasional reflection live on the notes page.