Reading and Steering Representations of Materials-Science Mechanisms in an Open-Weight Language Model
The paper argues that Gemma’s materials-science knowledge shows up more clearly in controlled hidden-state changes than in static representations.
Markus J. Buehler tests three open-weight Gemma 4 models on whether they encode mechanisms behind materials-science answers, not just the answers themselves. The study reports readable concept signals, hidden-state neighborhoods by mechanism, and directional state shifts that matched physical laws in 39 of 40 directional cases. It also finds one apparent organization could be explained by numerical comparison, so the paper leans on counterfactual prompts and causal interventions instead. HF Daily Papers' note
Markus J. Buehler tests three open-weight Gemma 4 models on whether they encode mechanisms behind materials-science answers, not just the answers themselves. The study reports readable concept signals, hidden-state neighborhoods by mechanism, and directional state shifts that matched physical laws in 39 of 40 directional cases. It also finds one apparent organization could be explained by numerical comparison, so the paper leans on counterfactual prompts and causal interventions instead. HF Daily Papers' note
score 4