What Constrains Adaptation After Pretraining? Generalization and Specialization Under Inherited Data Manifolds

TMLR Paper7035 Authors

Abstract: Large language models are often adapted to new tasks through supervised fine tuning. In deployment, however, their generalization can be unreliable and hard to anticipate. We examine whether such failures arise from limitations in optimization and supervision, or from geometric constraints inherited from pretraining, noting that data organization in representation space is rarely treated as an explicit control variable. Using controlled sampling from a large text distribution drawn from the web, we treat training samples as structured populations in representation space. We then compare data drawn from central and peripheral regions of the inherited manifold under identical architectures and training procedures. We find that data location in representation space strongly constrains what can be learned, frequently necessitating specialization across both general and domain-specific settings. Models trained on data drawn from peripheral or highly overlapping regions tend to generalize poorly, even when the training setup is otherwise unchanged. This pattern points to the need for principled specialization to meet practical demands on reliability, efficiency, and deployment.

Submission Type: Regular submission (no more than 12 pages of main content)

Assigned Action Editor: ~Adin_Ramirez_Rivera1

Submission Number: 7035