Ita-Lib at SVELA: Detecting the Forgotten - Representation-Based Approach for Verifying Machine Unlearning

Machine unlearning aims to remove specific knowledge from trained models while preserving their overall capabilities, a requirement that is increasingly relevant for legal, ethical, and safety reasons. While several unlearning techniques have been proposed for large language models (LLMs), reliably verifying whether a model has selectively forgotten targeted information remains an open challenge. In this paper, we present our submission to the SVELA shared task at EVALITA 2026, which focuses on the selective verification of erasure from LLM outputs. Our approach leverages model-derived representations to build a lightweight classifier for detecting forgotten knowledge in a model-agnostic, post-hoc setting that does not require access to training data or unlearning procedures. We evaluate the method on the multilingual dataset provided by the task and report results on the official benchmark. These findings suggest that while representation-based verification is promising, performance is constrained by the inherent challenges of the benchmark, and further research is needed to reliably detect erased knowledge across model scales and languages.