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

MIT researcher Markus J. Buehler investigates whether a language model’s correct materials-science answer reflects physical reasoning or merely lexical and numerical shortcuts. Using Google’s open-weight google/gemma-4-E4B-it, a 42-layer transformer, the study combines direct unembedding, three independently fitted Jacobian lenses, hidden-state geometry, counterfactual comparisons, and causal activation interventions. In 50 held-out descriptions spanning ten mechanism families, target-free decoded word sets enabled blinded identification of 9 of 10 families, although direct and Jacobian readouts performed similarly and prompt wording remained a strong confound. A 72-prompt graph showed comparative organization at the natural question boundary, with AUC 0.642, but exact audits found that its apparent physical polarity was equally explained by numerical direction. The strongest representational result came from a neutral-anchored 60-law benchmark: matched hidden-state changes distinguished direct, inverse, and neutral constitutive relations, correctly orienting 39 of 40 directional laws, while lexical controls were near chance. Causal steering was narrower but compelling: one grain-size direction reversed appropriately between refinement and coarsening in all 12 of 12 matched conditions, yet failed when transferred to new answer vocabularies and the inverse Hall–Petch regime. Full-state patching transferred late numerical and decision features across mechanisms, but not a universal physical-relation representation. An OpenAI gpt-5.5 interpreter recognized 9 of 10 mechanism families from decoded word sets, reinforcing semantic readability without proving causal use. The paper’s central conclusion is that materials relationships are more reliably revealed by controlled state transformations than by absolute hidden-state geometry, and that causal interpretability remains context-dependent.