Steering Materials Science Concepts in LLMs for Engineering.

Markus J. Buehler· July 23, 2026 View original

Summary

Researchers demonstrated that an open-weight language model (Google Gemma-4B-it) internally represents materials science mechanisms in three separable forms: readable concepts in hidden states, constitutive orientation in state transformations, and causal control over engineering answers. This allows for steering the model's physical understanding.

Large language models (LLMs) can answer complex scientific questions, but it's often unclear if they truly "understand" the underlying physics or merely mimic patterns. This research investigates how materials science mechanisms are represented within an open-weight LLM, specifically Google's Gemma-4B-it model. The findings reveal that mechanism information exists in three distinct forms within the model's internal states. Firstly, specific materials science concepts are directly readable within individual hidden states. Secondly, the model's internal state transformations carry constitutive orientation, meaning the direction of physical laws is encoded in how states change. Thirdly, selected internal representations can causally control the model's engineering-related answers, indicating a deeper, actionable understanding. Through a combination of direct and Jacobian vocabulary readouts, state geometry analysis, a counterfactual benchmark, and causal interventions, the researchers were able to identify and even manipulate these internal representations. This ability to "steer" the model's physical understanding suggests a path towards more reliable and controllable scientific AI applications.

Why it matters

This research provides crucial insights into how LLMs encode scientific knowledge, paving the way for developing more trustworthy and steerable AI tools for scientific discovery and engineering design.

How to implement this in your domain

  1. 1Explore techniques for interpreting and visualizing internal representations of LLMs relevant to domain-specific knowledge.
  2. 2Develop methods for fine-tuning or prompting LLMs to align their internal representations with specific scientific principles.
  3. 3Design experiments to test the causal influence of internal model states on desired scientific or engineering outputs.
  4. 4Collaborate with AI researchers to apply "steering" techniques to improve the reliability of LLM-generated scientific hypotheses or designs.

Who benefits

Materials ScienceChemical EngineeringPharmaceuticalsAerospace

Key takeaways

  • LLMs encode materials science mechanisms in readable hidden states, state transformations, and causal controls.
  • Concepts are readable in individual hidden states.
  • Constitutive orientation is carried by controlled transformations between states.
  • Internal representations can causally control engineering answers, enabling "steering."

Original post by Markus J. Buehler

"arXiv:2607.20058v1 Announce Type: new 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 we show that materials science mechanism information in the open-weight goo…"

View on X

Originally posted by Markus J. Buehler on X · view source

Want to go deeper?

Turn these trends into skills with Learnijoy's hands-on AI & tech courses.

Explore courses