
Spencer Kroll MD PhD
Our Scientific Focus
Verent organizes its work into four interconnected research domains

01
Alignment and Safety
Developing theoretical and empirical tools to ensure AI systems behave as intended under novel conditions, including adversarial and edge-case scenarios.
02
Interpretable Machine Learning
Creating techniques that expose the internal logic of complex models, enabling meaningful oversight by humans with and without technical backgrounds.
03
Distributed Intelligence Systems
Designing multi-agent frameworks that distribute cognitive load and risk across decentralized networks without sacrificing coordination or accountability.
04
Applied Reasoning
Building AI systems that perform structured inference across domains requiring precision, such as scientific hypothesis generation, legal reasoning, and medical decision support.
Projects Currently Under Investigation
A selection of active research programs at Verent
01
Project Calmere
_2026-09-09_12-48-23%201.jpg)
A two-year study into the behavioral stability of large language models under distributional shift. Partners include two European universities and one independent policy institute.
02
Project Varlind
_0010%208.jpg)
Development of a modular interpretability toolkit designed for use by non-technical regulatory auditors. Currently in beta evaluation with three national partner bodies.
03
Project Dusken

A theoretical investigation into privacy-preserving coordination protocols for networks of semi-autonomous AI agents operating in sensitive institutional environments.
Selected Publications
Verent researchers publish in leading peer-reviewed venues across AI, cognitive science, and ethics
Boundary Conditions for Behavioral Alignment in Autoregressive Systems
Halvare, O. and Coudreau, S. Published in the Journal of Artificial Intelligence Research. This paper proposes a formal taxonomy of alignment failure modes and introduces a testable constraint framework.

Auditable Attention: A Layer-Wise Transparency Protocol for Transformer Architectures

Coudreau, S. and Felmari, T. Published in the Proceedings of the International Conference on Machine Learning. Introduces a post-hoc interpretability method with near-zero inference overhead.
Decentralized Coordination under Epistemic Uncertainty in Multi-Agent Environments
Halvare, O. and Coudreau, S. Published in the Journal of Artificial Intelligence Research. This paper proposes a formal taxonomy of alignment failure modes and introduces a testable constraint framework.
