LLMs implicitly identify informative context in early layers, enabling a linear prober to drop noisy chunks and improve both performance and efficiency across tasks and models.
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
UNVERDICTED 1representative citing papers
citing papers explorer
-
END: Early Noise Dropping for Efficient and Effective Context Denoising
LLMs implicitly identify informative context in early layers, enabling a linear prober to drop noisy chunks and improve both performance and efficiency across tasks and models.