MIT reads and steers LLMs' internal physics intuition via hidden-state shifts in Gemma
ProfBuehlerMIT · x · 2026-11-19
MIT's Markus Buehler published new work in AI for Science showing that an LLM's representation of physical relationships can be both read and steered, with implications for post-training signals that deepen generalization of physical principles.
The method treats hidden states as measurable objects: across three sizes of Google's Gemma, the authors flipped a single physical quantity in a question (like pulling on a spring) and measured the resulting hidden-state shift. Two mathematical lenses interpret those numbers: the direct logit lens (which words the model favors at that stage) and the Jacobian lens (how sensitive downstream numbers are to small upstream changes). When describing grooves forming along crystal boundaries in steel, the Jacobian lens ranked corrosion first—although the question never named it—revealing the model's inner reasoning.
Key findings:
- Direction encodes causality: the direction of the internal shift correctly predicted whether increasing a quantity makes a material stiffer or softer. The pattern persisted when equations were rearranged or replaced with plain-language descriptions; both larger models captured the correct direction for all laws predicting a change.
- Similarity graph: connecting questions whose internal number patterns look alike groups problems within a mechanism simply because both describe an increase.
- Internal intervention: strengthening only the grain-size signal in the model's internal numbers made it favor higher strength for smaller grains across six distinct metals, with the question text untouched—evidence these representations causally support predictions.
Implication: physical laws live inside LLMs in a human-readable, steerable form, and such post-training signals could help models build deeper generalizations of physical principles.
Related event: MIT Researchers Read and Manipulate LLMs' Internal Physics Intuition(2 posts)→
More from Research
- Sandia Labs' VASt model predicts high-entropy alloy properties from tens of samples — CurieuxExplorer · 2026-11-19
- FoodTruck Bench, an LLM agent business-sim benchmark, lands on Papers with Code — NielsRogge · 2026-11-19
- harness-hub open-sources RL-trained LoRA adapters tuned per agent harness, recipe coming Friday — yb2698 · 2026-11-19
- Freenet releases new video on network routing with old-school ML — Sanity · 2026-11-19
- Empire AI creates 120+ early-career research positions in New York, Cornell deadline Dec 15 — dianarycai · 2026-11-19
- Author fires back at ICLR withdrawal backlash: rebuttal got zero responses — hbouammar · 2026-11-19