cs.LGNov 8, 2026

Full Fine-Tuning is not the Problem: Why Adam Fails and SGD Succeeds in Few-Shot CLIP Adaptation

Authors: Nairouz Mrabah, Ismail Ben Ayed, Eric Granger

Organizations: LIVIA, ILLS, Department of Systems Engineering, ETS Montreal, Canada

Abstract

Few-shot CLIP adaptation is often treated as a parameterization problem. Full fine-tuning is assumed to overfit, so parameter-efficient fine-tuning (PEFT) methods are preferred. In this paper, we show that this conclusion conflates parameterization with optimization. Across 11 datasets, 3 CLIP backbones, and 5 shot levels, dense full fine-tuning with plain SGD outperforms representative PEFT methods under the reported evaluation protocol, whereas Adam/AdamW collapses. We trace this collapse to unreliable support-set preconditioning. The few-shot second moment systematically underestimates the corresponding full-data statistic, already at initialization. Adam's inverse-square-root normalization converts this error into an oversized adaptive update. Matching Adam's first-step update budget to SGD resolves most of the collapse, whereas numerator-side alignment does not explain the failure. Shampoo and SOAP also fail, so richer preconditioning alone does not resolve the problem. Theory formalizes how few-shot estimation error is amplified into unsafe updates. SGD succeeds because it avoids unreliable preconditioners and stays close to pretrained CLIP. Interpolation probes show low-loss connections along tested directions, and curvature probes show favorable directional curvature. These results identify optimizer-induced preconditioning, not dense full fine-tuning itself, as the main source of failure in few-shot CLIP adaptation.

Figures & tables

Appendix figures & tables133 assets

Supplementary material from the paper’s appendix.

Appendix

Explore similar work

Oct 26, 2026

LE-GDA: Layer Ensemble Gaussian Discriminant Analysis for Training-Free Few-Shot CLIP Adaptation

Training-free few-shot adaptation of vision-language models typically constructs a classifier from CLIP's final projected image representation, leaving intermediate representations unused. Combining them is not straightforward. Class scores from different representations differ in spread by orders of magnitude for the same query. Direct averaging allocates influence by numerical scale rather than predictive information. We propose Layer-Ensemble Gaussian Discriminant Analysis (LE-GDA), a method that fits regularized shared-covariance discriminants independently to intermediate and final representations extracted in one frozen encoder pass. LE-GDA standardizes each discriminant's class scores per query, selects a small representation multiset on labeled validation data, and combines it with the zero-shot classifier without gradient-based optimization. Across 11 datasets, four CLIP backbones, five shot counts, and three independently sampled K-shot training sets, LE-GDA achieves the highest mean accuracy of all evaluated methods on every backbone, raising the 16-shot mean from 78.01% to 79.54% over a matched GDA-CLIP (final-feature) baseline and improving that baseline for every backbone-shot combination. At one shot, LE-GDA retains a 1.20-point mean gain over the baseline even though covariance estimation is unavailable and per-query standardization is neutral. A control with richer per-class structure inside a single representation does not reproduce the gain, indicating that the benefit comes from combining representations. These results show that intermediate CLIP representations carry information beyond the final embedding and can improve training-free few-shot adaptation. Code and evaluation resources are available at https://github.com/pavodi-NM/le-gda.
Mar 4, 2026

Hold-One-Shot-Out (HOSO) for Validation-Free Few-Shot CLIP Adapters

In many CLIP adaptation methods, a blending ratio hyperparameter controls the trade-off between general pretrained CLIP knowledge and the limited, dataset-specific supervision from the few-shot cases. Most few-shot CLIP adaptation techniques report results by ablation of the blending ratio on the test set or require additional validation sets to select the blending ratio per dataset, and thus are not strictly few-shot. We present a simple, validation-free method for learning the blending ratio in CLIP adaptation. Hold-One-Shot-Out (HOSO) presents a novel approach for CLIP-Adapter-style methods to compete in the newly established validation-free setting. CLIP-Adapter with HOSO (HOSO-Adapter) learns the blending ratio using a one-shot, hold-out set, while the adapter trains on the remaining few-shot support examples. Under the validation-free few-shot protocol, HOSO-Adapter outperforms the CLIP-Adapter baseline by more than 4 percentage points on average across 11 standard few-shot datasets. Interestingly, in the 8- and 16-shot settings, HOSO-Adapter outperforms CLIP-Adapter even with the optimal blending ratio selected on the test set. Ablation studies validate the use of a one-shot hold-out mechanism, decoupled training, and improvements over the naively learnt blending ratio baseline. Code is released here: https://github.com/chris-vorster/HOSO-Adapter
Nov 5, 2026

On the Few-Shot Adaptation of Vision-Language Models: A Few Gradient Signs Are All You Need

The limited supervision available for the few-shot adaptation of vision-language models makes gradient estimation particularly noisy. We find that the gradient magnitudes are severely miscalibrated, while the signs of a small top-magnitude subset remain informative. This motivates SignSparseSGD and DynSignSparseSGD, two state-free optimizers that use gradient magnitudes only to select where to update and fixed-magnitude signs to decide how to update. The static variant fixes the support after the first minibatch. The dynamic variant refreshes it at every step. Across 11 datasets, three CLIP backbones, and five shot levels, the proposed methods are competitive with or outperform parameter-efficient and optimization-based baselines on average, while exhibiting markedly lower sensitivity to long training schedules. They introduce no task-specific modules, maintain no first- or second-moment optimizer states, and use one shared configuration across datasets, backbones, and shot levels. Our code is included in the supplementary materials and will be released upon publication.