In gradient-free weight perturbation of LLMs, the perturbation norm controls success; a frozen frame of about a dozen scalars nearly matches full-weight search, and a fixed scale of 0.05 works across seven models.
In NeurIPS Datasets and Benchmarks
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.CL 1years
2026 1verdicts
CONDITIONAL 1representative citing papers
citing papers explorer
-
Not the Dimension, the Norm: What Matters in Gradient-Free Weight Perturbation of Language Models
In gradient-free weight perturbation of LLMs, the perturbation norm controls success; a frozen frame of about a dozen scalars nearly matches full-weight search, and a fixed scale of 0.05 works across seven models.