A new preprint on arXiv introduces a distinction between two kinds of facts a language model can recall. Active facts support generalization beyond their own recall, while inert facts can be recalled but do not support such further learning. The authors argue that not all recalled knowledge is equally useful for downstream tasks.

The paper's title points to a key finding: knowledge that is linear before fine-tuning tends to generalize after it. This suggests that the geometric structure of a fact's representation—specifically its linearity—may indicate whether that fact will support learning beyond simple retrieval.

The abstract is truncated, so the full methodology and experimental details are not available from the source. Still, the framing implies a practical direction: identifying which facts are likely to transfer to new tasks before fine-tuning, rather than after the fact.