Reading and Steering Representations of Materials-Science Mechanisms in an Open-Weight Language Model
Year: 2026
Type: article
Abstract: Large language models can answer scientific questions, yet a correct output does not reveal whether the model represents or uses the governing physics. Here, using three open-weight Gemma 4 instruction-tuned models (google/gemma-4-E4B-it, google/gemma-4-12B-it, and google/gemma-4-31B-it) we identify three experimentally separable signatures of materials-science mechanism information: selective concept readability, relational encoding of qualitative constitutive orientation, and causal, context-dependent control of constrained engineering answers. Across 50 held-out descriptions, three independently fitted Jacobian lenses produced nearly identical concept ranks, but 36 prompts recovered no declared term within the top 100 and the family-level advantage over direct unembedding was not significant. Target-free words from both readouts nevertheless enabled blinded automated identification of 9 of 10 mechanism families. A separate 72-prompt benchmark formed mechanism-specific hidden-state neighborhoods, but an exact graph audit showed that numerical comparison explained the same organization; absolute state geometry therefore did not uniquely identify physics. In a matched 60-law benchmark, state changes caused solely by reversing a physical input ordered direct, physically neutral and inverse relations ($\rho=0.910$) and correctly oriented 39 of 40 directional laws. Removing the equation and prescribed reasoning scaffold preserved correct orientation for 38 of 40 laws at 4B and all 40 laws at both 12B and 31B. A prospectively frozen grain-size direction steered all 12 refinement and coarsening cases toward the correct answer. In a disjoint 24-law cohort, neutralizing, reversing and restoring a development-fitted relational coordinate selectively altered answer evidence across all three scales, while activation-matched controls remained near zero. Qualitative physical relationships are therefore revealed more clearly by controlled changes in representation than by absolute state alone, and selected relational representations can causally influence scientific decisions.
Source: AI for Science
Author: Markus J. Buehler
Institution: Massachusetts Institute of Technology
Language: English
Subfield: Materials Chemistry
Field: Materials Science
Domain: Physical Sciences
Open Access status: diamond
OpenAlex ID: W7170177592