PLOT: Progressive Localization via Optimal Transport in Neural Causal Abstraction
Mechanistic interpretability, which seeks to uncover the internal computations underlying neural model behavior, is central for AI safety by helping us diagnose, anticipate, and steer frontier AI systems. Causal abstraction offers a principled framework for mechanistic interpretability, aligning a high-level causal model with low-level neural computation through interchange intervention analysis. Finding such an alignment, however, is a challenging task that often requires fitting and evaluating separate learned mappings across many candidate neural locations.
I will give an overview of neural causal abstraction and introduce PLOT, a gradient-free approach that automates correspondence discovery through global optimal transport (OT) matching. PLOT represents abstract and neural interventions by geometric signatures of their effects on the task output and fits an OT coupling between them. This coupling can be calibrated into an executable intervention handle or used progressively to guide finer localization. Examples from hierarchical equality, binary addition, and multiple-choice question answering will illustrate the method and its performance, with accurate handles and substantial computational gains. Time permitting, I will discuss ongoing research on using PLOT to select faithful causal models, gradual handle discovery, and localization over interpretable features rather than directly over neural activations.
Bio: Ziv Goldfeld is an associate professor in the School of Electrical and Computer Engineering at Cornell University, and a graduate field member in computer science, statistics and data science, operations research and information engineering, and the Center for Applied Mathematics. Ziv’s research interests include optimal transport theory, information theory, statistics, optimization, and AI alignment and interpretability. He develops the mathematical and algorithmic foundations of machine learning and AI algorithms, subject to formal guarantees on reliability, efficiency, scalability, robustness, and safety. Prior to joining Cornell, he was a postdoctoral fellow in the Laboratory for Information and Decision Systems (LIDS) at MIT. He received his B.Sc., M.Sc. (both summa cum laude), and Ph.D. in electrical and computer engineering from Ben-Gurion University of the Negev. Honors include the NSF CAREER Award, the IBM University Award, the Rothschild postdoctoral fellowship, and the Michael Tien ’72 Excellence in Teaching Award at Cornell.