BEGIN:VCALENDAR VERSION:2.0 PRODID:-//Date iCal//NONSGML kigkonsult.se iCalcreator 2.20.2// METHOD:PUBLISH X-WR-CALNAME;VALUE=TEXT:Eventi DIAG BEGIN:VTIMEZONE TZID:Europe/Paris BEGIN:STANDARD DTSTART:20251026T030000 TZOFFSETFROM:+0200 TZOFFSETTO:+0100 RDATE:20261025T030000 TZNAME:CET END:STANDARD BEGIN:DAYLIGHT DTSTART:20260329T020000 TZOFFSETFROM:+0100 TZOFFSETTO:+0200 TZNAME:CEST END:DAYLIGHT END:VTIMEZONE BEGIN:VEVENT UID:calendar.30789.field_data.0@www.diag.uniroma1.it DTSTAMP:20260518T113717Z CREATED:20260506T044347Z DESCRIPTION:AbstractSafety Neurons in Large Language ModelsLarge language m odels (LLMs) achieve state-of-the-art performance across a wide range of t asks\, but their widespread deployment raises urgent concerns around secur ity\, privacy\, and misuse. Building on recent progress in sparse mechanis tic interpretability\, particularly findings from vision models\, this tal k examines the hypothesis that a small set of neurons or features may play a disproportionate role in safety-aligned behavior in LLMs. We begin by p resenting methods for identifying such sparse\, interpretable substructure s and evaluating how inference-time manipulation of these components can d egrade safety behavior in both white-box and black-box settings. We then e xtend this perspective to Mixture-of-Experts (MoE) models\, introducing a training-free\, lightweight\, and architecture-agnostic framework for prob ing and stress-testing the safety alignment of modern MoE LLMs during infe rence. Finally\, the talk discusses the broader implications and applicati ons of “safety features\,” including their role in safety-relevant behavio r in code-generation models\, and highlights opportunities for more robust alignment and defense.Bio. Stjepan Picek is a full professor at the Unive rsity of Zagreb\, Faculty of Electrical Engineering and Computing\, Croati a. He also holds an associate professor position at Radboud University\, N ijmegen\, and an adjunct professor position at the University of Bergen\, Norway. Before that\, he was an assistant professor at TU Delft and a post doctoral researcher at MIT\, USA\, and KU Leuven\, Belgium. Stjepan comple ted PhD in computer science in 2015 at the University of Zagreb\, Croatia and Radboud University\, The Netherlands. In 2024\, he finished a PhD in m athematics at the University of Paris 8\, France. His research interests i nclude security and cryptography\, machine learning\, and evolutionary com putation. To date\, Stjepan has given more than 80 invited talks and publi shed more than 200 refereed papers. He is a program committee member and r eviewer for a number of conferences and journals and a member of several p rofessional societies. His work has been featured in the mainstream media and on popular technology blogs. He is a member of ELLIS and a Fellow of t he Young Academy of Europe.Stjepan Picek is visiting Sapienza/DIAG in the context of the EMAI program\, in order to conduct research and teaching ac tivities. DTSTART;TZID=Europe/Paris:20260508T100000 DTEND;TZID=Europe/Paris:20260508T100000 LAST-MODIFIED:20260506T055427Z LOCATION:Aula A3\, via Ariosto 25\, Roma SUMMARY:Safety Neurons in Large Language Models - Stjepan Picek URL;TYPE=URI:https://www.diag.uniroma1.it/node/30789 END:VEVENT END:VCALENDAR