Online SFT data selection acts as an implicit preference model, shifting refusal rates, verbosity, and sycophancy in directions predictable from the selected data's attribute mixture.
Learning without forgetting.IEEE Transactions on Pattern Analysis and Machine Intelligence, 2017
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.LG 1years
2026 1verdicts
CONDITIONAL 1representative citing papers
citing papers explorer
-
Online Data Selection Is Implicit Alignment
Online SFT data selection acts as an implicit preference model, shifting refusal rates, verbosity, and sycophancy in directions predictable from the selected data's attribute mixture.