Reading and Steering Representations of Materials-Science Mechanisms in an Open-Weight Language Model

This paper investigates whether the open-weight google/gemma-4-E4B-it model represents materials-science mechanisms rather than merely producing plausible scientific language. Using vocabulary readouts, hidden-state geometry, a 60-law counterfactual benchmark, and causal interventions, the authors find that mechanism concepts are readable from individual states, while constitutive orientation is more reliably encoded in controlled state transformations. The transformations correctly oriented 39 of 40 directional laws, and bidirectional interventions shifted answer probabilities toward or away from physically appropriate outcomes across 12 matched cases.