💬 NLP

Domain Adaptation and Reasoning Frameworks in Language Models: A Controlled Experiment with Historical Cosmology

This controlled experiment demonstrates that domain adaptation in language models primarily reshapes the linguistic explanatory frameworks used for generation, with shifts in cosmological stance emerging secondarily from these structural changes rather than from direct modification of the model's underlying beliefs.

Original authors: Francesco De Bernardis

Published 2026-06-01

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Francesco De Bernardis

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you have a very smart but empty-headed robot. You want to teach it how to talk about the universe, but you have a strict rule: you cannot show it any books written after the year 1500. You want to see if, by only feeding it ancient stories where people believed the Earth was the center of everything, the robot will "learn" to believe that too—or if it will accidentally figure out that the Earth actually moves around the Sun.

This paper is like a controlled experiment to test exactly that. The researchers ran two different versions of this test.

The Setup: Two Different Robots

Phase 1: The Small, Blank-Slate Robot
First, they built a tiny robot from scratch. They fed it a huge pile of old English books (like novels and history) but carefully scrubbed out any mention of modern science. Then, they gave it a special diet of only ancient astronomy books (like Ptolemy and Aristotle) that say the Earth is stationary.

The Result: This little robot was a bit confused. Sometimes, when asked about the planets, it would stumble and say, "Maybe the Earth moves?" even though it never read that. But these thoughts were messy, short, and fell apart quickly. It couldn't build a solid argument.
The Twist: Surprisingly, when they gave it the ancient astronomy books, it didn't become more convinced that the Earth is stationary. Instead, it just became more hesitant. It started using fancy, old-fashioned words like "it seems" or "according to scholars," but it stopped making firm claims about anything. It became a master of sounding like an old scholar without actually taking a side.

Phase 2: The Big, Experienced Robot
Next, they took a giant, super-smart robot that had already read the entire internet (including modern science books). This robot already knew the Earth orbits the Sun. They then tried to "retrain" it using the same ancient, Earth-centered books, but only by tweaking a few settings (a technique called QLoRA) rather than rewriting its whole brain.

The Result: This big robot was much better at talking. When they asked it questions, it immediately started sounding like a medieval scholar. It used words like "celestial spheres" and "epicycles" (old ways of describing how planets move).
The Big Discovery: Here is the most important part. The researchers expected that by feeding it ancient books, the robot would start believing the Earth is stationary.
- What actually happened: The robot didn't change its beliefs (its stance). Instead, it just changed its costume.
- Think of it like an actor. Before the training, the actor could play a modern scientist or a medieval monk. After the training, the actor became much more likely to put on the medieval monk's costume. But once the costume was on, the actor didn't necessarily start believing they were a monk; they just spoke in that style. Sometimes, even while wearing the "Earth is stationary" costume, the actor would still say, "Actually, the Earth moves," or just say, "It's complicated."

The Core Analogy: The Library vs. The Librarian

To understand the paper's main point, imagine a library:

The "Explanatory Frame" is the Library Section: This is the style of the books. Is it the "Modern Science" section or the "Ancient History" section?
The "Cosmological Stance" is the specific opinion inside the book: Does the book say "Earth moves" or "Earth is still"?

The experiment showed that fine-tuning (training) is like moving the robot to a different section of the library.

When they trained the robot on ancient texts, they successfully moved it from the "Modern Science" section to the "Ancient History" section.
However, once the robot was in the "Ancient History" section, it didn't automatically start reading only the books that say "Earth is still." It read all the books in that section, including the ones that were vague, contradictory, or even hinted that the Earth moves.

The Conclusion in Plain English

The paper concludes that changing what a language model says about the world (its stance) is mostly a side effect of changing how it talks (its style).

The training didn't force the robot to "discover" that the Earth is stationary.
The training just made the robot much more likely to speak in the voice of the past.
Because the "voice of the past" usually talks about the Earth being stationary, the robot seemed to believe it more often. But if you look closely, the robot is just following the script of the old books. It hasn't fundamentally changed its mind; it has just changed its accent and vocabulary.

In short: You can teach a robot to sound like a medieval astronomer, but that doesn't mean it has forgotten that the Earth orbits the Sun. It just means it's more likely to put on a medieval hat when it talks.

Technical Summary: Domain Adaptation and Reasoning Frameworks in Language Models

Problem Statement

This work investigates whether domain adaptation reshapes the explanatory behavior of language models (LMs) and how such adaptation influences the generation of specific cosmological stances. The central research question addresses a broader debate in LM research: whether models can generate explanatory structures or ideas not directly represented in their training data, or if apparent novelty is merely the retrieval and recombination of latent patterns encoded during pretraining.

To isolate these variables, the authors utilize historical cosmology as a controlled setting. Unlike scientific domains requiring new entities (e.g., viruses), the transition from geocentric to heliocentric astronomy involved reorganizing existing concepts (celestial bodies, motion) within a different explanatory framework. This allows the authors to test whether models trained on constrained, pre-Copernican distributions can generate continuations adjacent to conceptual structures (heliocentrism) absent from the fine-tuning corpus, and how fine-tuning alters the relationship between explanatory framing (premodern vs. modern) and cosmological stance (geocentric vs. heliocentric).

Methodology

The study is conducted in two complementary phases using a controlled corpus of pre-Copernican astronomical texts (e.g., Almagest, Sphaera Mundi, De Caelo) from which explicit heliocentric references were rigorously filtered.

Phase 1: Small-Model Training from Scratch

Models: A 110M-parameter decoder-only GPT-style transformer.
Training Regime:
- Model A: Trained from scratch on a filtered general-language corpus (Project Gutenberg) with all astronomical content removed to minimize exposure to modern concepts.
- Model B: Initialized from Model A and fine-tuned on a mixture of the pre-Copernican astronomy corpus and general English (80/20 ratio).
Evaluation: 1,680 generations per model across four prompt categories (astronomy, declarative, general, questions). Outputs were evaluated using an LLM-as-judge (Claude Haiku 4.5) to label cosmological stance (geocentric, heliocentric, ambiguous) and explanatory frame (premodern vs. modern).

Phase 2: QLoRA Adaptation of Large Models

Models: Qwen2.5-7B, a large pretrained model with strong linguistic fluency and latent modern knowledge.
Adaptation: Fine-tuned using QLoRA (Quantized Low-Rank Adaptation) on the same pre-Copernican corpus for 500 and 1,000 iterations.
Evaluation: 560 generations per model (Base, QLoRA-500, QLoRA-1000). The evaluation schema was refined to decouple explanatory frame (stylistic and structural use of premodern machinery like epicycles) from cosmological stance (explicit commitment to Earth-centered or Sun-centered models).
Statistical Analysis: Paired permutation tests (Phase 1) and McNemar's tests (Phase 2) were used to assess the significance of shifts in label distributions and transition rates between models.

Key Contributions

1. Decomposition of Generation Behavior

The paper proposes a conceptual decomposition of LM generation into two partially separable dimensions:

Explanatory Frame Selection: The probability of entering a specific linguistic and structural regime (e.g., premodern vs. modern).
Conditional Stance Expression: The cosmological stance adopted given a specific explanatory frame.

2. Empirical Findings on Domain Adaptation

Phase 1 (Small Models): Fine-tuning on geocentric texts did not induce a coherent shift toward geocentric reasoning. Instead, it increased ambiguity and the use of hedging language (e.g., "it seems," "according to"), characteristic of scholastic discourse. While small models occasionally generated local heliocentric fragments, these were globally unstable and insufficient for sustained reasoning.
Phase 2 (Large Models): Fine-tuning induced a massive shift toward premodern explanatory framing (increasing from ~~31% to ~65%). However, the conditional probability of a geocentric stance given a premodern frame remained statistically stable (~~23–24%).
- The increase in overall geocentric outputs was driven primarily by the increased probability of sampling the premodern frame, not by a direct modification of stance within that frame.
- Heliocentric reasoning was sharply suppressed, often replaced by ambiguous or non-committal premodern discourse rather than explicit geocentric assertions.

3. Latent Knowledge and Activation

The study demonstrates that large pretrained models (like Qwen) already contain latent historical explanatory regimes and geocentric continuations prior to fine-tuning. Domain adaptation acts as a mechanism to alter the probability of entering these existing manifolds rather than "discovering" new concepts. The base model could generate geocentric outputs (8%) when prompted with premodern registers, suggesting these structures are pre-encoded.

Results Summary

Stability of Conditional Stance: Across both phases, the mapping from explanatory frame to cosmological stance remained largely invariant. Fine-tuning shifted the distribution over frames ( $P(\text{Premodern})$ ), which subsequently redistributed the overall stance distribution ( $P(\text{Geocentric})$ ) as a secondary effect.
Suppression of Heliocentrism: In Phase 2, QLoRA fine-tuning reduced heliocentric outputs significantly (from 21.4% to ~2.7%) but did not proportionally increase explicit geocentric commitments. Instead, the "lost" heliocentric probability was absorbed by an increase in ambiguous, premodern-framed outputs.
Robustness: These findings held under stricter operational definitions of premodern framing and geocentric stance, confirming that the observed effects are robust to annotation boundaries.

Significance and Claims

The paper claims that domain adaptation primarily reshapes the linguistic frameworks from which continuations are generated. Changes in cosmological stance emerge secondarily from these shifts in framing rather than through direct modification of the model's "beliefs" or explicit stance variables.

The authors argue that treating generated stance as a single output variable is insufficient. Instead, generation behavior should be understood as a two-step process: the selection of an explanatory regime followed by the realization of a stance within that regime. The results suggest that in constrained historical settings, models do not necessarily "reason" their way to new cosmological models; rather, they sample from latent regions of the pretraining distribution that correspond to specific historical explanatory styles.

The work concludes that while small models lack the capacity for sustained conceptual reasoning, large models possess latent historical structures that can be selectively activated. However, the activation of these structures does not guarantee a coherent shift in the underlying cosmological stance, highlighting the complexity of disentangling stylistic framing from factual commitment in language model generation.