InfoSFT modifies the standard SFT loss with a one-line change to weight tokens by information content, targeting medium-likelihood ones for better generalization and capability retention on math, code, and reasoning tasks.
Title resolution pending
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
-
InfoSFT: Learn More and Forget Less with Information-Aware Token Weighting
InfoSFT modifies the standard SFT loss with a one-line change to weight tokens by information content, targeting medium-likelihood ones for better generalization and capability retention on math, code, and reasoning tasks.