Ander Artola Velasco

Hi! I’m Ander, a PhD. student at the Max Planck Institute for Software Systems, under the supervision of Dr. Manuel Gomez Rodriguez and as part of the IMPRS-TRUST program. Previously, I obtained my masters in Physics at Heidelberg University under the supervision of Dr. Sarah Bosman. I graduated in 2023 from the Complutense University of Madrid in Mathematics and Theoretical Physics.

I'm currently seeking research internships in industry – focused on decision-making and strategic behavior in LLMs, inference, optimization, uncertainty quantification and evaluation. If my research interests align with your team’s goals, I’d be excited to connect!

Some of my (past) interests include:

Decision-making and strategic behavior in machine learning systems, including large language models
Bayesian modelling, inference and optimization
Game-theoretic statistics
Uncertainty quantification, explainability and support systems
Causality
(Differential geometry)
(Cosmology, dark matter, the intergalactic medium and reionisation)

news

May 11, 2026	We have a new preprint! Many products (including new medical treatments!) are heavily subsidized by regulators during development and approval. But developers respond strategically to those subsidies, and a poorly designed one can waste resources or kill the very moonshots it was meant to encourage. We study a sequential approval protocol where the regulator partially subsidizes randomized controlled trials and the developer chooses, round by round, whether to keep experimenting or walk away. We characterize the developer’s optimal strategy, and show how the regulator can subsidize them to maximize social utility. On real antibiotic data, the protocol yields >35% more social utility than one-shot randomized controlled trials.
Apr 30, 2026	Our paper “Is Your LLM Overcharging You? Tokenization, Transparency, and Incentives” has been accepted as a Spotlight paper at ICML’26!
Jan 30, 2026	We have a new preprint! In this paper we introduce test-time compute games, a game-theoretic framework for analyzing how LLM providers choose the amount of test-time compute deployed by their models. Building on this framework, we show that current LLM-as-a-service markets are socially inefficient, and we propose a forward-looking auction mechanism that is provably socially optimal.
Jan 30, 2026	Our LLM auditing paper was accepted at AISTATS’26!
Nov 15, 2025	I will give a spotlight talk at the PAIG workshop at EurIPS’25!