Speaker: Chenhao Zhang

Logistics: Wednesday May 6th 12-1pm in Mudd 3514. Lunch will be provided.

Title: Barriers to Counterfactual Credit Attribution for Autoregressive Models

Abstract: Generative AI disrupts the practice of giving credit to work that came before. Ideally, a generative model would give credit to any work on which its output depends in a significant way. In this talk, I will present my recent joint work with Aloni Cohen from the University of Chicago on the theory of credit attribution of generative models. We consider the proper credit attribution notion counterfactual credit attribution (CCA)---a relaxation of differential privacy---recently introduced by Livni, Moran, Nissim, and Pabbaraju (NeurIPS 2024) for classification models. We initiate the study of CCA autoregressive generative (AR) models and uncover barriers to two natural approaches to obtaining them. First, an AR model's next-token predictor satisfying CCA does not automatically imply that the model itself satisfies CCA. Second, adding non-excessive credit to an existing non-credit-attribution AR model to satisfy CCA is computationally hard.

Bio: Chenhao Zhang is a final-year Computer Science PhD student working with Jason Hartline and Christos Dimoulas. His primary research lies in the intersection of computer science and law using various tools spanning from the theory of algorithms to programming languages.

Note: This talk is open to everyone, including faculty members as well as friends outside the CS community. Critical feedback from those working on practical aspects of LLM are particularly welcomed (regardless of being constructive or having fundamental disagreement on the limitation of theoretical research).