Training on LLM text embeddings can cause tabular classifiers to collapse to single-class predictions, which spuriously inflates Accuracy-on-the-Line correlations.
The column ”Well-defined” refers to empirical observation that the corresponding HP sets do not collapse for the tabular modality in our setting
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.LG 1years
2025 1verdicts
CONDITIONAL 1representative citing papers
citing papers explorer
-
Data Curation Matters: Model Collapse and Spurious Shift Performance Prediction from Training on Uncurated Text Embeddings
Training on LLM text embeddings can cause tabular classifiers to collapse to single-class predictions, which spuriously inflates Accuracy-on-the-Line correlations.