GRASS adaptively samples LLM layers using gradient norms and offloads optimizer states to achieve up to 4.38 points higher accuracy and 19.97% lower memory than prior fine-tuning methods.
InInternational Conference on Learning Representations
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.CL 1years
2026 1verdicts
UNVERDICTED 1representative citing papers
citing papers explorer
-
GRASS: Gradient-based Adaptive Layer-wise Importance Sampling for Memory-efficient Large Language Model Fine-tuning
GRASS adaptively samples LLM layers using gradient norms and offloads optimizer states to achieve up to 4.38 points higher accuracy and 19.97% lower memory than prior fine-tuning methods.