The Big Picture 2026:
Crafting a Research Narrative v2

Where it Hurts: Finding Durable Questions While Moving Fast

In a fast-moving field, it can be hard to tell which problems are urgent, which are merely loud, and which are worth building a research life around. This talk considers how researchers can stay responsive to rapid change without letting the field’s volatility set their agenda. I will discuss tools for identifying which questions remain meaningful across shifts in methods, data, benchmarks, and institutions.  The goal is a practical vocabulary for finding direction when everything seems to be moving at once.

Bio  Noah A. Smith is the inaugural Vice Provost for Artificial Intelligence and Charles and Lisa Simonyi Endowed Chair for Artificial Intelligence and Emerging Technologies at the University of Washington, where he is also a Professor in the Paul G. Allen School of Computer Science & Engineering. He is Senior Director of NLP Research at the Allen Institute for Artificial Intelligence, directs the OLMo open language modeling effort, and leads the NSF- and NVIDIA-supported project “Open Multimodal AI Infrastructure to Accelerate Science.” His research spans language and music technologies, multimodal AI, and multifaceted evaluation of AI systems.

Does mechanistic interpretability need interventions?

Mechanistic interpretability often treats interventions as the gold standard of evidence, relying on circuit ablations and representation steering to support claims about how models actually work. But are interventions really sufficient, or even necessary, for making mechanistic claims? In this debate-style talk, we trace the history that led the field to embrace interventions, and argue that the answer to both questions is a contentious "no." We start by showing that interventions alone are not sufficient to explain model behavior by highlighting cases where causal methods can produce misleading or outright spurious explanations. Then, we debate whether interventions are necessary at all, exploring how alternative notions of causality and carefully designed behavioral evidence may also support strong mechanistic claims without directly intervening on a model. We conclude by discussing future directions for mechanistic interpretability, and how we can draw inspiration from other scientific disciplines to ask what should count as a good explanation.

Bios  Tiago Pimentel is a Postdoctoral Researcher at ETH Zürich, working in machine learning interpretability and psycholinguistics. His long-term goal is to understand how humans and machines process language. To this end, his research adopts an interdisciplinary approach, leveraging information theory and causality to study the mechanisms behind model behaviour and human cognition.

Aaron Mueller is an Assistant Professor of Computer Science and, by courtesy, of Data Science at Boston University. His research centers on developing interpretability and evaluation methods inspired by causal and linguistic principles, and applying these to precisely control and improve the generalization of language technologies. He completed his Ph.D. at Johns Hopkins University.