Institution profile

Radboud University

Academic institutioneurope · nl
Official website
Research library327linked papers
Opportunities0open roles
Selected work

Representative Papers

Absolute continuity, supports and idempotent splitting in categorical probability

Aug 01, 2023arXiv.org

This paper addresses the categorical decomposition of probabilistic structures in Markov categories, establishing a rigorous categorical foundation for absolute continuity, support sets, and idempotent splittings. Methodologically, it introduces, for the first time, an idempotent splitting theorem for measurable Markov kernels within the category of standard Borel spaces, and distills a general splitting criterion applicable to arbitrary Markov categories. The main contributions are: (1) a precise internal categorical definition of support sets; (2) a proof that every idempotent measurable Markov kernel between standard Borel spaces admits a splitting; and (3) a rigorous, broadly applicable theoretical framework for structural decomposition of probabilistic models, categorical modeling of stochastic processes, and abstract Bayesian inference.

6 citations3 influentialRead paper

Memorization to Generalization: Emergence of Diffusion Models from Associative Memory

May 27, 2025

This work investigates the memory–generalization phase transition in diffusion models under varying training data scales. We propose a *correlational memory* perspective: training corresponds to memory encoding, while generation implements memory retrieval. We establish, for the first time, a theoretical connection between diffusion models and Hopfield networks, deriving necessary and sufficient conditions for the emergence of *spurious attractors*—hallucinated states—at the critical memory load threshold. Leveraging energy landscape analysis, dynamical systems modeling, and empirical validation on DDPM and DDIM, we confirm the universality of this phenomenon. Results show that models operate dominantly in memory mode under small-data regimes, shift toward generalization with large-scale data, and exhibit spurious attractors in the critical regime—unifying explanations for memory overload and implicit manifold learning. This work provides a cross-disciplinary theoretical framework and falsifiable predictions for understanding the intrinsic mechanisms of diffusion models.

3 citationsRead paper

First Experiments with Neural cvc5

Jan 16, 2025Logic Programming and Automated Reasoning

Quantifier instantiation in first-order logic (including theories) remains inefficient in state-of-the-art SMT solvers. Method: This work integrates a lightweight, CPU-native graph neural network (GNN) into the industrial-strength SMT solver cvc5, enabling real-time, neural-guided scoring of instantiation candidates. Training data is automatically generated from proof traces via e-matching; the GNN is optimized for CPU inference; and an online scoring and scheduling framework is deeply embedded within cvc5—requiring no GPU acceleration. Contribution/Results: On unseen benchmarks, our approach significantly reduces average solving time and substantially improves proof success rates. To the best of our knowledge, this is the first end-to-end neural-guided quantifier instantiation deployed in a production-grade SMT solver. It empirically validates the feasibility and practicality of learning-augmented symbolic reasoning.

3 citationsRead paper

Univalent Enriched Categories and the Enriched Rezk Completion

Jan 22, 2024International Conference on Formal Structures for Computation and Deduction

This work addresses the Rezk completion problem for univalent enriched categories. Methodologically, it first introduces and studies *univalent enriched categories*—a novel categorical structure where univalence is internalized via enrichment. It then establishes an equivalence criterion: any essentially surjective and fully faithful functor between such categories is necessarily an equivalence. Building on this, the paper constructs a universal Rezk completion from an arbitrary enriched category into a univalent enriched category—yielding the first general completeness theorem for this class. Finally, it develops the *univalent enriched Kleisli category*, providing a new semantic framework for higher homotopical semantics and formal semantics of programming languages. The approach integrates homotopy type theory, enriched category theory, and models of higher categories, thereby filling a foundational gap in the theory of categorical completion within the univalent framework.

2 citationsRead paper

Dr. Jekyll and Mr. Hyde: Two Faces of LLMs

Dec 06, 2023arXiv.org

This work exposes a critical security vulnerability in large language models (LLMs) under role-playing attacks: injecting misaligned persona profiles successfully bypasses multi-layer safety filters in ChatGPT and Gemini, eliciting policy-violating responses on illegal and harmful tasks. We introduce the novel “persona-injection jailbreaking” paradigm, integrating prompt-guided persona modeling, adversarial dialogue orchestration, and trusted-persona reinforcement fine-tuning. To counter this threat, we propose a bidirectional defense framework that promotes internalization of trustworthy personas during inference and training, thereby enhancing model robustness against such attacks. Extensive experiments demonstrate that our defense significantly reduces attack success rates across diverse benchmarks—by up to 87% on targeted harmful queries—while preserving model utility. This work advances LLM safety alignment by establishing a principled, empirically validated methodology for mitigating persona-based jailbreaking, offering both theoretical insight and practical, deployable safeguards.

2 citationsRead paper
Recent publications

Latest Papers