MIT's Buehler reads and steers materials-science physics representations inside open-weight Gemma 4 LLMs

ProfBuehlerMIT · x · 2026-11-19

MIT professor Markus J. Buehler published an open-access paper in IOP's new journal AI for Science, probing whether LLMs genuinely represent physics rather than just produce correct answers. Using three open-weight Gemma 4 instruction-tuned models (4B/12B/31B), the study identifies three experimentally separable signatures of materials-science mechanism information: selective concept readability, relational encoding of qualitative constitutive orientation (whether increasing one quantity makes another rise, fall or stay unchanged), and causal context-dependent control of constrained engineering answers. The direction of internal representational shifts reveals causal orientation between physical quantities, suggesting post-training signals help models build deeper generalizations of physical principles — and these representations can be steered, not just read.

Related event: MIT Researchers Read and Manipulate LLMs' Internal Physics Intuition(2 posts)→

Original post →

More from Research

Research channel →