A gradient-guided regularization method for few-shot LLM fine-tuning reports higher average accuracy than baselines on SuperGLUE, but lacks code, error bars, and a practical optimization recipe.
Learning from models beyond fine-tuning,
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.CL 1years
2025 1verdicts
REJECT 1representative citing papers
citing papers explorer
-
Structured Gradient Guidance for Few-Shot Adaptation in Large Language Models
A gradient-guided regularization method for few-shot LLM fine-tuning reports higher average accuracy than baselines on SuperGLUE, but lacks code, error bars, and a practical optimization recipe.