IC-1106CLIP prompts found by greedy search do not fit random labels: train and test error drop in tandem as the fraction of flipped labels increases, unlike a linear probe which achieves near-random accuracy

Victor Akinwande, Yiding Jiang, Dylan Sam, J Zico Kolter

SourceUnderstanding prompt engineering may not require rethinking generalization

The authors test whether CLIP prompts can memorize noise by running greedy search on CIFAR-10, CIFAR-100, FMOW, and OfficeHome with a controlled fraction of randomly flipped training labels. Both training and test error decrease monotonically and in tandem as the flip ratio increases, indicating the prompts track the true signal rather than the noise. As a baseline, a linear probe trained on the same CLIP features with 100% flipped labels achieves roughly random performance (13.60% accuracy on CIFAR-10), confirming that the prompt space is too small to memorize arbitrary label assignments.

Evidence
correlational
Key metric
Linear probe on CIFAR-10 with 100% flipped labels achieves 13.60% accuracy; greedy prompts show monotonic in-tandem drop of train and test error across 0-100% label flip ratio on CIFAR-10, CIFAR-100, FMOW, and OfficeHome
Caveat
The experiment uses a fixed prompt length of 5 and the CLIP vocabulary; the result is specific to the discrete prompt formulation and may not extend to continuous (soft) prompts.
Model
CLIP / CLIP-ViT (LC)
Datasets
CIFAR-10 [eval], CIFAR-100 [eval], FMOW [eval], OfficeHome [eval]
Methods
Linear Probing / Ridge regression linear probing / Linear probe / Linear probe fine-tuning / Linear regression probing / Linear ridge regression probes / Supervised probing / ERM linear probe [compared-to]
Related findings
IC-1105
Extraction
automatic-extraction