Researchers from the Faculty have achieved an impressive milestone, with three papers from Dr. Hagai Maron’s research group accepted to ICLR 2026, NeurIPS 2025, and AAAI 2026, among the world’s most prestigious conferences in machine learning and artificial intelligence.
Conducted in collaboration with researchers from additional universities and NVIDIA, the work addresses one of the most pressing challenges in modern AI: identifying when large language models (LLMs) generate inaccurate information, exhibit biases, or deviate from their intended behavior.
The research was led by Ph.D. student Guy Bar-Shalom (co-supervised by Prof. Ran El-Yaniv) and postdoctoral researcher Dr. Fabrizio Frasca, in collaboration with Dr. Yiftach Ziser (University of Groningen and NVIDIA).
The team introduces a novel approach for detecting failures, risks, and hallucinations in large language models. Rather than attempting to fully interpret the complex internal mechanisms of these systems, the researchers developed an efficient and scalable method that leverages the models’ internal computations to monitor, diagnose, and predict problematic behavior. The approach enables external oversight and control, even without complete understanding of the model’s training process or internal architecture.
The new methodology opens the door to a wide range of applications in areas where reliability is essential, including healthcare, education, research, regulation, and safety-critical systems. It may support the development of warning systems, quality assurance tools, and safety standards for the responsible deployment of AI technologies.
These papers are part of a broader research program in Dr. Maron’s laboratory, which explores how new forms of information can be extracted from trained models to improve the understanding, reliability, and safety of advanced AI systems.
Photo (right to left): Dr. Fabrizio Frasca, Dr. Hagai Maron, and Guy Bar-Shalom.
