AAAM Research

LLM Bias Measurement

As Large Language Models become increasingly central to how people access and process information, concerns about their tendency to introduce and amplify social biases, including ethnic, cultural, nationality, age, and socioeconomic prejudices, have grown significantly. Existing benchmarks typically evaluate bias using static, offline datasets, leaving users of live LLM-based systems without immediate, interpretable feedback on the fairness of generated content.

This research activity focuses on a demographic anchor-based method to approximate bias in LLM outputs for unknown, real-world content. The core idea is to measure how closely a generated text aligns with pre-defined demographic anchors, reference texts associated with specific social groups, adapted from the BBQ dataset, using cosine similarity in an embedding space. By projecting both the generated content and the demographic anchors into a shared semantic space via a local embedding pipeline (all-MiniLM-L6-v2), the method yields a quantitative bias score without requiring labelled ground-truth data. This makes it applicable to arbitrary, previously unseen outputs, across bias categories including complex intersectional cases.

Applied Category Theory

The aim is to abstract concepts and notions of artificial intelligence research and applications and capture them in terms of category theory, in order to reason better about them.

This research activity is driven by Luiza Corpaci organizing a (bi-)weekly reading club and discussion forum on Mondays 18:00 - 19:30 (Central european Time) (UTC+1) at Google Meet. For more information, do not hesitate to contact Luiza directly (luiza@[OurAcronym].info).

Combinatorial Methods for Test Vector Generation in Black-Box Hardware Security Testing

This collaborative research activity between Dr. Fatma Nur Esirci Oral and members of the AAAM was initiated during Fatma Nur’s research visit from September 4 to October 4, 2025.

The activity is dedicated to the exploration of combinatorial methods for test vector generation in the context of black-box security testing of hardware.

The planned work, which combines Dr. Esirci Oral’s expertise with that of AAAM members, aims to compare different testing strategies, provide solid benchmark results, and develop resources to help the wider community address security challenges in digital circuits.