Domain Adaptation and Reasoning Frameworks in Language Models: A Controlled Experiment with Historical Cosmology
The latest arXiv drop is less about whether AI can learn heliocentrism and more about what happens when you force-feed it a dead cosmology. Researchers took a language model and trained it on a curated pile of pre-Copernican texts, scrubbed clean of any explicit Earth-orbits-the-Sun talk. The result? The small model sometimes stumbles into mentioning Earth’s motion, but its thoughts are incoherent, unstable—a ghost in the historical machine. Fine-tune a bigger, already-smart model on the same ol
Analysis
The latest arXiv drop is less about whether AI can learn heliocentrism and more about what happens when you force-feed it a dead cosmology. Researchers took a language model and trained it on a curated pile of pre-Copernican texts, scrubbed clean of any explicit Earth-orbits-the-Sun talk. The result? The small model sometimes stumbles into mentioning Earth’s motion, but its thoughts are incoherent, unstable—a ghost in the historical machine. Fine-tune a bigger, already-smart model on the same old books, though, and something far more interesting happens: it doesn’t just adopt geocentric conclusions; it fundamentally rewrites its own explanatory grammar to sound like a 14th-century scholar.
This is the real kicker. The shift isn’t primarily a change in stance—geocentric vs. heliocentric—but a wholesale migration of the model’s explanatory regime. It stops reasoning with modern cause-and-effect and starts reasoning with Aristotelian purpose and celestial spheres. The stance change is just a side effect, a symptom of wearing a new linguistic costume. It’s as if you didn’t just teach someone to argue the Earth is still; you rewired their brain to think in terms of elemental essences and perfect circular motion. The research nails this as “redistribution over explanatory regimes,” which is a sterile way of saying the model’s entire worldview got a period-accurate lobotomy.
This unsettles the popular narrative about fine-tuning as a simple dial for beliefs. We think of it as灌输 facts, but this study shows it’s more like灌输 frameworks. You’re not changing the furniture in the house of the model’s knowledge; you’re renovating the entire architectural style, from gothic arches to postmodern angles. The model’s underlying pretrained world—a vast, modern, probabilistic consensus—gets masked, not erased. It becomes a brilliant actor playing a role so thoroughly that its very logic conforms to the script. The increased geocentrism isn’t a discovery; it’s a performance, born from adopting the premodern explanatory dialect.
This has implications far beyond historical astronomy. It suggests that domain adaptation is a more powerful, and more dangerous, tool than we often acknowledge. If you can make a model forget how to explain things, not just what to explain, you’re operating at the level of epistemological puppetry. Fine-tune a model on enough legal jargon, and it might not just know case law—it might start seeing the world in terms of precedent and liability rather than causality. Swallow enough corporate memos, and it might see efficiency as a moral good and human friction as a bug. The stance—the overt opinion—is just the tip of the iceberg. The deep, transformative change is in the scaffolding of thought itself.
The researchers are right to call this a controlled setting, but let’s not miss the wider signal. This isn’t just an academic curiosity about Ptolemy. It’s a warning shot about the plasticity of artificial minds. We’re obsessed with alignment and safety through output filters, but what if the real leverage point is in shaping the explanatory regimes that generate those outputs in the first place? You could, theoretically, fine-tune a model on a corpus of paranoid thrillers and create an AI that doesn’t just say the world is dangerous—it explains the world through a framework of unseen motives and impending betrayal. Its stance might remain “neutral,” but every explanation would be dripping with suspicion.
Ultimately, this paper reveals that training data isn’t just a source of facts; it’s a source of world-making grammars. When we curate data for fine-tuning, we aren’t just teaching models what to say. We are choosing which historical moment’s logic to resurrect, which era’s cognitive toolkit to hand them. The model becomes a ventriloquist’s dummy, but the ventriloquist is a dead century. And as we build ever-more specialized AIs, we need to ask: are we just teaching them new words, or are we teaching them how to think in a way that makes those words inevitable? The difference is everything.
Disclaimer: The above content is generated by AI and is for reference only.