Papers/2608.23634
🧪 Test?View on arXiv

The Blending Ratio Is Not Where the Performance Is: Diagnosing Prototype Blending for Few-Shot Adaptation of Vision-Language Models

Liangzhi Li, Author 2, Author 3, Author 4, Author 5

few-shot learningvision-language modelsprototype blendingvalidation-free methods
2608.23634
Builder Relevance
80%
3h ago

Abstract

This paper investigates the effectiveness of blending ratios in few-shot adaptation of vision-language models, revealing that optimal ratios do not correlate with performance improvements.

Reality Card

Core Claim

The study demonstrates that the theoretically optimal blending ratio does not lead to the best performance, and validation-free methods can achieve better results.

Method / Result

Validation-free linear probes outperform the oracle-tuned blend by +1.9 points on average.

Limitations

The reliance on specific datasets and models may limit the generalizability of the findings.

Paper to code

Verified implementation resources so builders can test the paper’s claims instead of stopping at the abstract.

No verified implementation link has been attached yet. AIBuzzHub will keep this panel separate from unverified search results.
← Back to all papers