Divergent Tasks Harm Integration Of New Entities Via Fine-Tuning

Core Francisco Park

Keywords: Fine-Tuning, Representation, Symbol Grounding, Multi Task Learning

TL;DR: Some tasks are "divergent" and fine-tuning on them harms cross-task generalization by encoding new entities in hidden subspaces.

Abstract: When a pretrained model is fine-tuned to incorporate new entities, does this knowledge automatically generalize across tasks? We investigate this question using a controlled framework in which transformers pretrained on geometric tasks over real-world city coordinates are fine-tuned to integrate synthetic out-of-distribution cities. We find that cross-task generalization varies dramatically depending on the fine-tuning task, and that this variation is partially predicted by representational similarity (CKA) measured during pretraining. In multi-task fine-tuning, we identify lurking divergent tasks that not only fail to generalize but actively harm performance on other tasks. Probing suggests that divergent tasks encode new entities in separate subspaces rather than integrating them into the shared world manifold.

Anonymization: This submission has been anonymized for double-blind review via the removal of identifying information such as names, affiliations, and identifying URLs.

Style Files: I have used the style files.

Submission Number: 92