AI Paper Atlas

Research map / Architectures and efficiency / Pruning and efficient tuning

The Geometry of Narrow Fine-Tuning Degradation: Trajectory Lock-in and Spectral Bifurcation

Jia Liu, Jiaxin Luo, Xinhao Qiu, Yixue Hao, Min Chen

ICML 2026 · Architectures and efficiency › Pruning and efficient tuning · approach: Adapter-based parameter-efficient tuning

Read the paperCheck your idea against 137,217 papers

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
add SVD spectral regularizer to LoRA fine-tuningBaseline LoRA · Narrow LoRA fine-tuningfinal broad loss: 4.54 -> 3.42Improved
add SVD spectral regularizer to LoRA fine-tuningBaseline LoRA · Narrow LoRA fine-tuningmean Abs-Harm: 0.020 -> 0.012Improved
add SVD spectral regularizer to LoRA fine-tuningBaseline LoRA · Narrow LoRA fine-tuning regime comparisonfinal broad loss: 5.10 -> 3.51Improved
add SVD spectral regularizer to LoRA fine-tuningBaseline LoRA · Narrow LoRA fine-tuning regime comparisonfinal task loss: 0.28 -> 0.31No clear effect
add gradient noise injection to LoRA fine-tuningBaseline LoRA · Narrow LoRA fine-tuning regime comparisonfinal broad loss: 5.10 -> 2.99Improved
add gradient noise injection to LoRA fine-tuningBaseline LoRA · Narrow LoRA fine-tuning regime comparisonfinal task loss: 0.28 -> 0.97Hurt
add SVD regularizer overhead vs baselineBaseline LoRA · Training overheadtime per step (ms): 328.7 -> 350.4Hurt
add null-space projection to LoRA fine-tuningBaseline LoRA · Training overheadtime per step (ms): 328.7 -> 382.1Hurt
increase SVD regularization strength s (weak to moderate to very strong) (ablation)SVD regularizer · Narrow LoRA fine-tuningbroad loss at matched task lossMixed
delay SVD rescue to later in training (ablation)SVD regularizer · Narrow LoRA fine-tuningprotection rateHurt
apply SVD regularization in micro-dataset full fine-tuning (N=64)SVD regularizer · Full fine-tuning N=64broad lossNo clear effect

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy