FIM-LoRA allocates per-layer ranks in LoRA using calibration-time gradient variance as a proxy for informativeness via an efficient eFIM diagonal approximation, matching standard LoRA accuracy on GLUE and commonsense tasks while yielding interpretable rank patterns.
Title resolution pending
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.LG 1years
2026 1verdicts
UNVERDICTED 1representative citing papers
citing papers explorer
-
FIM-LoRA: Task-Informative Rank Allocation for LoRA via Calibration-Time Gradient-Variance Estimation
FIM-LoRA allocates per-layer ranks in LoRA using calibration-time gradient variance as a proxy for informativeness via an efficient eFIM diagonal approximation, matching standard LoRA accuracy on GLUE and commonsense tasks while yielding interpretable rank patterns.