Parameter-efficient fine-tuning (LoRA family)
SNGP
Superseded baseline#211 of 1,113 most-superseded
Superseded — cited as a baseline and beaten by newer methods
1 papers critique it · 1 beat it on benchmarks
What papers say
Verbatim critique sentences, each from a paper that cites SNGP as a baseline.
However, SNGP's bi-Lipschitz assumptions do not hold for transformers, as dot-product self-attention has an unbounded Lipschitz constant.
Beaten on benchmarks
Head-to-head results where a newer method reports beating SNGP. Values are copied from the source paper's tables — verify against the cited paper.
LoRA-Ensemble beats SNGP
82.5 vs 32.2
Accuracy · [16-member ensemble on CIFAR-100]
LoRA-Ensemble: Efficient Uncertainty Modelling for Self-Attention Networks