Ofra Amir
Technion

Abstract: Policies learned by reinforcement learning (RL) agents are often hard to explain, emerging from complex reward structures and opaque neural representations. Existing global summarization methods use action demonstrations in select states, but users can only watch so much—and must infer patterns on their own. In this talk, I present SySLLM, a method that uses large language models (LLMs) to synthesize behavioral patterns into concise textual summaries, leveraging their world knowledge and pattern recognition. Time permitting, I’ll also share recent work on self-consistency in LLMs—how well their explanations align with their actual reasoning.

Short bio: Ofra Amir is an Associate Professor at the Faculty of Data and Decision Sciences at the Technion. She completed her B.Sc. and M.Sc. in Information Systems Engineering at Ben-Gurion University and holds a Ph.D. in Computer Science from Harvard University. Her research interests are at the intersection of AI and human-computer interaction.

Michael Fire
Ben-Gurion University

Abstract: Large Language Models (LLMs) transform research, industry, and everyday life. Yet, their scale and flexibility introduce a new class of security threats. LLMs face a critical vulnerability: jailbreak attacks that bypass safety mechanisms to generate harmful content. This talk presents our research on "dark LLMs," deliberately unaligned models and universal jailbreak techniques that compromise state-of-the-art commercial systems.

Our research reveals that a publicly known jailbreak method remains effective against leading LLMs despite being available for months. Industry responses to our responsible disclosure efforts were inadequate, highlighting gaps in AI safety practices. We also examine the growing ecosystem of unaligned models. We propose comprehensive defense strategies including training data curation, LLM Guardrails, machine unlearning, and continuous red teaming. Unlike centrally managed platforms, open-source vulnerabilities cannot be patched once released, making immediate action critical. Without decisive intervention, we risk democratizing access to dangerous knowledge at an unprecedented scale. This talk demonstrates the urgent threats and practical solutions for making LLMs safer.

Short bio: Dr. Michael Fire is a Senior Lecturer (Assistant Professor) in the Faculty of Computer and Information Science at Ben-Gurion University (BGU) and the founder of the Fire AI Lab. He was awarded the prestigious Moore/Sloan Data Science Fellowship and the WRF Innovation Postdoctoral Fellowship in Data Science at the University of Washington. With over 60 published papers, Michael's research focuses on Safe AI, Applied AI, Big Data, and Security and Privacy. He collaborates with researchers and organizations to develop AI-driven solutions for real-world, multidisciplinary challenges.

Idan Schwartz
Bar-Ilan University

Abstract: In recent years, the exponential growth in data and model sizes has led to significant advancements in generative language and image models. Despite these improvements, generative models often require large scale fine-tuning to adapt to specific domains or tasks. This fine-tuning typically involves a curated dataset and an additional, computationally intensive learning phase. In this talk, I will briefly show alternatives. First, when data is available, I will show an encoder approach to turn a text-to-video model into an audio-to-video model by training an audio-to-text mapping. Next, I will show an inference-time approach with no data that adds control to video generation by allowing temporal conditioning, for example when an object appears or an action takes place, by optimizing cross-attention maps. Lastly, I will discuss an approach that employs an external model during inference time to enhance class accuracy, object counting, and subject-driven generation.

Short bio: Idan Schwartz is an Assistant Professor at Bar-Ilan University, where he leads the Multimodal Lab. His research centers on multimodal learning, with a focus on generative models, modal alignment, and efficient inference-time solutions. He received his PhD in Computer Science from the Technion under Prof. Tamir Hazan and Prof. Alexander G. Schwing (UIUC), and completed a postdoc with Prof. Lior Wolf at Tel Aviv University. His industry experience includes research roles at eBay, Microsoft, and Spot on vision–language systems such as catalogs and dialogs, and on time-series prediction.