27cc3576a6f149e95cf68afc3e25cd6c.zip Today

While the reviews were generally positive, experts noted a few areas for improvement:

Reviewers highlighted that the paper's design choices, specifically "feature sharing," were well-motivated and helped the model stay expressive despite the simplifications. Critical Perspectives 27cc3576a6f149e95cf68afc3e25cd6c.zip

This paper introduces a method called designed to improve how we tune large "black-box" models (like CLIP) when we don't have access to their internal code or gradients. Performance and Efficiency While the reviews were generally positive, experts noted

The primary consensus among reviewers is that ZIP significantly reduces the "query cost"—the number of times you have to ask the model for a result—while maintaining or improving accuracy. While the reviews were generally positive

Evaluators noted superior accuracy across 13+ different tasks and strong performance in "few-shot" settings (learning from very little data).

It addresses the high query requirements of existing methods by reducing problem dimensionality and using "intrinsic-dimensional gradient clipping."

One reviewer pointed out that the methods ZIP was compared against (like BLACKVIP and BPTVLM) were from 2023, and suggested that more recent 2024 benchmarks should have been included for a fairer comparison.