Proxy loss functions can outperform the target loss when training AI assessors, with logistic loss and logarithmic score being the most promising proxies in the reported experiments.
Title resolution pending
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
-
What should an AI assessor optimise for?
Proxy loss functions can outperform the target loss when training AI assessors, with logistic loss and logarithmic score being the most promising proxies in the reported experiments.