Research Fellow Candidate Presentations

Please hold this time for presentations from our 2025-2026 Research Fellow candidates.

9:00-9:45am – Hadas Orgad, Israel Institute of Technology

Talk Title: Leveraging Interpretability for AI Safety and Robustness

• As AI models like ChatGPT continue to advance in versatility and performance, they also bring critical challenges in robustness and safety that can be difficult to address in diverse, real-world use cases. Interpretability research, which aims to elucidate the decision-making process of AI models, offers a promising pathway to address these challenges with customizable and cost-effective methods. In this webinar, I will present our research on enhancing AI robustness and safety by applying insights from interpretability studies, focusing on mitigating biases, reducing harmful content, improving adaptability, and addressing hallucinations.

 9:45-10:30am – Gizem Ozdil, EPFL

Talk Title: Reverse-engineering the neural mechanism underlying multi-body part coordination in Drosophila

• Complex animal behaviors arise from the coordination of multiple body parts, yet the neural mechanisms underlying multi-body part coordination remain poorly understood. In this talk, I will present a combined experimental and computational approach to study multi-body part coordination in the fruit fly, Drosophila melanogaster. I will show that motor coordination during antennal grooming is primarily driven by a central mechanism composed of distinct network motifs rather than sensory feedback.

10:30-11:15am – TBD

11:15am-12:00pm – William Merrill, New York University

Talk Title: The Parallelism Tradeoff: Understanding the Power and Limitations of Transformers Through Circuit Complexity

• Scaling transformer language models has enabled tremendous progress in NLP and deep learning, but how far can this paradigm be pushed? In this talk, I will discuss my body of theoretical results on the theoretical expressive power of transformers and related architectures, and how these results bear on this question. I will discuss our theoretical result that transformers (without chain of thought) can only express problems in the complexity class uniform TC0 and thus cannot express many simple computational problems including state tracking, compositional formula evaluation, and graph connectivity. I will then discuss our work characterizing how chain of thought approaches can expand the expressive power of transformers, as well as our work comparing the expressive power of state-space models and transformers. Overall, these findings reveal a fundamental tradeoff between parallelism and expressive power: the parallelism so essential for scaling up transformer language models also precludes them from expressing many simple computational problems. These insights let us more precisely understand the limitations of transformers and also provide a strong foundation upon which to develop novel language modeling architectures and inference methods.

Moderator: Cengiz Pehlevan