Scientific Causal Reasoning for Interpretability –

Published

September 12, 2026

Summary

We’re doing a series of experiments in order to look into whether we can use scientific assays and constructions in order to understand if models are causally using internal representations of scientific concepts in order to solve problems. We’re doing this through using primarily the OpenADMET PXR dataset, and running both mech interp experiments and behavioural experiments, on scientific concepts like selectivity, potency, and dose-dependent responses.

AI usage disclosure incomplete. Before submission or release, complete ai-use.yml. Every first and last author must describe their exact AI usage, and every remaining author must read the complete disclosure and confirm it in that file.

Purpose

We want to share experiments as we go along, including our null results, imperfect designs, and how we iterate through prompt design.

Learn more

Please reach out to us, or feel free to look through our blog posts!