The authors finetuned six vision-language models from several model families and scales on six class-imbalanced image-classification datasets.
Different inference prompts recover rare classes suppressed during VLM finetuning
Across six datasets and six vision-language models, switching to a different descriptive prompt at evaluation restored rare-class recall without retraining.
Big Tech
Emanuel Ben Baruch · Sivan Biham · Alon Shoshan · Yochai Zvik · Avi Ben-Cohen
Amazon
Research Digest··2 min read
Ben Baruch and colleagues study vision-language models finetuned on datasets where a few classes dominate and others have scarce examples.
Why this paper
From Amazon
In one line
Using a different descriptive inference prompt than the finetuning prompt restores rare-class recall suppressed by imbalanced VLM finetuning.
What we could check
- ·No code link found
- ·No weights link found
- ·No dataset link found
- ·No compute details found
- ·No stated limitations found
- ·No benchmark numbers found
Observed from the paper text and links we have. Absence here means we did not find it, not that it does not exist.
§