Post

Florian Eichin
@florian-eichin.com

๐—š๐—ฟ๐—ฎ๐—ฑ๐—ถ๐—ฒ๐—ป๐˜ ๐—ฑ๐—ฒ๐˜€๐—ฐ๐—ฒ๐—ป๐˜ ๐—ถ๐˜€ ๐—ฎ๐—น๐—น ๐˜†๐—ผ๐˜‚ ๐—ป๐—ฒ๐—ฒ๐—ฑ ๐—ณ๐—ผ๐—ฟ ๐—ถ๐—ป๐˜๐—ฒ๐—ฟ๐—ฝ๐—ฟ๐—ฒ๐˜๐—ฎ๐—ฏ๐—ถ๐—น๐—ถ๐˜๐˜†? In our ICML paper, we show that it might just be. ExPLAIND is a method that unifies data attribution, model component attribution, and training dynamics by computing an exact decomposition of model behavior based on gradient products.

1 repost6 likes

Florian Eichin
@florian-eichin.com

Want to learn more? Poster presentation at Tue, Jul 7, 2026, 2:00 PM โ€“ 3:45 PM @ HALL A #3303 Also in the workshop on mechanistic interpretability, Fri, Jul 10, 2026 โ€ข 8:00 AM โ€“ 5:00 PM @ HALL C Paper: arxiv.org/abs/2505.20076

arxiv.orgExPLAIND: Unifying Model, Data, and Training Attribution to Study Model BehaviorPost-hoc interpretability methods typically attribute a model's behavior to its components, data, or training trajectory in isolation, and are often tied to a particular level of granularity along the...