Department of Computer, Control and Management Engineering

Abstract

In this talk we present a approach to evaluating large language models (LLMs) based on the latest EU GPAI Code of Practice. It explores the mapping between existing technical benchmarks and the risk categories defined by the Code, linking model propensities, capabilities, and risk exposure. The intervention outlines how to operationalize safety measures required by the Code, including robustness tests, misuse evaluations, and systemic risk assessments. Emphasis is placed on translating abstract commitments into measurable technical processes to support compliance and align LLM development with EU regulatory expectations.