Preprint
Aug 2026
The Blending Ratio Is Not Where the Performance Is: Diagnosing Prototype Blending for Few-Shot Adaptation of Vision-Language Models
This work proves the mechanism and bound its share of the damage at 26% by a counterfactual, and finds the ceiling in the model class, not the hyperparameter: the ratio can be set near-optimally for free, and it is still not where the performance is.
Liangzhi Li, Bowen Wang, Yiming Qian et al.
· 0 citations