Is Laplace-LoRA superseded?

Q: Is Laplace-LoRA superseded?

Laplace-LoRA (Parameter-efficient fine-tuning (LoRA family)): cited as a baseline — critiqued by newer work, not yet beaten on a benchmark here. 2 paper(s) critique it, 0 beat it on benchmarks — #236 of 1113 most-superseded. Sub-problem: cluster led by MC Dropout.

Parameter-efficient fine-tuning (LoRA family)

Laplace-LoRA

Superseded baseline#236 of 1,113 most-superseded

Cited as a baseline — critiqued by newer work, not yet beaten on a benchmark here

2 papers critique it · 0 beat it on benchmarks

What papers say

Verbatim critique sentences, each from a paper that cites Laplace-LoRA as a baseline.

Laplace-LoRA is a post-hoc calibration method requiring longer training iterations to bring the low-rank parameters from an unstable basin (a subspace associated with the same local optimum) to a more stable parametric space. Therefore, Laplace-LoRA often leads to sub-optimal downstream performance.

— Robust and Efficient Fine-tuning of LLMs with Bayesian Reparameterization of Low-Rank Adaptation

“Methods such as Laplace-LoRA~yang2023bayesian require an additional pass through the data to compute a Hessian or Fisher approximation.”

— Variational Low-Rank Adaptation Using IVON