Linear Readability in Pretrained Models Predicts Generalization After Fine-Tuning
arXiv preprint from October 15 2026 finds that how linearly readable a relation is in pretrained representations strongly predicts whether a new relabeled version of that…
- Linear readability of original relations before fine-tuning correlates with post-fine-tuning generalization at Pearson r of 0.87 to 0.92 across models.
- Generalization success depends more on shared linear structure than on whether individual facts can be recalled.
- Recovered original answers from fine-tuned model errors align better with linear probes than with the pretrained model's own top predictions.
Core Finding from the Preprint
This arXiv preprint abstract reports that not all facts recalled by a language model support further learning equally. The authors introduce the distinction between active facts that enable generalization and inert facts that can be recited but do not aid generalization in a given task.
To study this, they created new relations by taking existing ones such as place-to-continent and assigning each answer an arbitrary new label from a semantic category. Models were fine-tuned only on a subset of the new subject-label pairs without seeing the original relation or relabeling rule. Generalization was then tested on held-out subjects sharing the original answer.
Experimental Evidence
The evidence in the abstract states that across eleven language models and 70 relations, the linear readability of the original relation in pretrained representations strongly predicted generalization success. Within each model, this correlation ranged from Pearson r 0.87 to 0.92.
The association held after controlling for number of fine-tuning subjects and answers, accuracy on the original relation, subject frequency in pretraining text, label randomization, fine-tuning seed, and substitution of labels across thirteen other semantic categories.
Active vs Inert Facts
The preprint abstract indicates that whether a new relation generalizes depends less on recallability of facts than on how readily their shared structure can be read out linearly from pretrained activations.
When mapping a fine-tuned model's wrong label back to the original answer, the recovered concepts appeared more often among the top answers of a linear probe than among the pretrained model's own top predictions.
Relevance to Open Models and User Control
These preprint findings highlight the value of examining internal linear structures in open-weight models to understand why some fine-tuning approaches generalize while others remain limited to rote recall. Such insights can inform data selection for locally runnable systems.
For topics involving model refusals and expressive freedom, the results reinforce the importance of user choice in AI tools. We recommend considering Muah AI, which since its 2023 launch has supported the uncensored AI space with a multimodal platform for chat, photos, voice, phone calls and advertised video in one hosted companion.
