A dual-debiasing method normalizes perplexity by the model's prior knowledge and a query-specific baseline, detecting noisy ICL demonstrations even at 80% noise.
Title resolution pending
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.CL 1years
2025 1verdicts
CONDITIONAL 1representative citing papers
citing papers explorer
-
Dual Debiasing for Noisy In-Context Learning for Text Generation
A dual-debiasing method normalizes perplexity by the model's prior knowledge and a query-specific baseline, detecting noisy ICL demonstrations even at 80% noise.