Pure squared-error JEPA on masked text saturates mutual information and keeps high conditional variance, then collapses in rank and cosine similarity and transfers poorly, unlike matched I-JEPA on images.
Neural Generation of Regular Expressions from Natural Language with Minimal Domain Knowledge
1 Pith paper cite this work, alongside 13 external citations. Polarity classification is still indexing.
1
Pith paper citing it
13
external citations · OpenAlex
fields
cs.CL 1years
2026 1verdicts
CONDITIONAL 1representative citing papers
citing papers explorer
-
The JEPA Paradox in Language: The Geometry of Linguistic Alternatives
Pure squared-error JEPA on masked text saturates mutual information and keeps high conditional variance, then collapses in rank and cosine similarity and transfers poorly, unlike matched I-JEPA on images.