Unilateral normalization variants outperform cosine and dot product on retrieval and other tasks when chosen according to whether the task treats query and candidate as functionally symmetric.
Removing normalization eliminates this projection, allowing gradients to flow more directly and potentially enabling faster convergence to better minima
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.LG 1years
2026 1verdicts
UNVERDICTED 1representative citing papers
citing papers explorer
-
When Does Embedding Magnitude Matter? A Cross-Task Functional-Symmetry Framework
Unilateral normalization variants outperform cosine and dot product on retrieval and other tasks when chosen according to whether the task treats query and candidate as functionally symmetric.