LLM fine-tuning on ancient texts shifts explanatory frames, not just stance

A new paper by Francesco De Bernardis explores how domain adaptation reshapes language model reasoning using historical cosmology as a controlled setting. In Phase 1, a small language model was trained from scratch on a pre-Copernican corpus with explicit heliocentric references removed. The model occasionally generated local Earth-motion continuations, but these were globally unstable and insufficient for coherent cosmological reasoning. In Phase 2, a larger pretrained model was fine-tuned using QLoRA on the same corpus to study how adaptation modifies explanatory framing and stance. Outputs were evaluated by an LLM-as-judge framework that labeled both cosmological stance (geocentric, heliocentric, ambiguous) and explanatory frame (premodern vs modern).

The results showed a large, statistically significant shift toward premodern explanatory framing after fine-tuning, while conditional stance distributions remained stable within those frames. The increase in geocentric outputs was primarily due to redistribution over explanatory regimes rather than direct stance modification. This suggests domain adaptation primarily reshapes the linguistic frameworks from which continuations are generated, with changes in stance emerging secondarily. The findings have implications for understanding how fine-tuning can alter not just surface outputs but the underlying reasoning structures of LLMs, particularly when dealing with historically or culturally specific datasets.