MIT Study Peers Inside LLMs: Internal Representations of Physical Laws Are Readable and Steerable
CatAstro_Piyush · x · 2026-11-20
MIT's Markus Buehler team published new research on the "internal metaphysics" of LLMs, appearing in AI for Science.
- Across a large set of materials-science laws, the direction of a shift inside the model revealed whether increasing one quantity makes another rise, fall, or stay the same
- The pattern persisted when equations were rearranged, and even when they were removed and described purely in words
- By nudging the model's internal activity, the team could test which relationships actually contributed to its predictions
- Key conclusion: LLMs represent physical relationships in a form that is both readable and steerable, offering insights into post-training signals that help models build deeper generalizations of physical principles
More from AGI Musings
- When AI explains better than humans, will we still learn from people? — danfaggella · 2026-11-20
- Prediction: by 2028 most patients will consult AI before seeing a doctor — realmeetjames · 2026-11-20
- Anthropic's 5,300 employees now out-revenue Tencent's 120,000 per head by 12x — 创业邦 · 2026-11-20
- Everful raises $48.5M to connect 30,000 Chinese factories with 70 staff using AI agents — 快鲤鱼 · 2026-11-20
- banteg's satire: DeepMind 'solves' recursive self-improvement with a schoolgirl persona — banteg · 2026-11-20
- Toby Ord clashes with Jan Leike over whether human control of AI is a consensus — tobyordoxford · 2026-11-20