Is Less More? Exploring Token Condensation as Training-free Test-time Adaptation

TL;DR

This paper introduces Token Condensation as Adaptation (TCA), a training-free method that improves vision-language model performance on unseen datasets by efficiently condensing tokens, achieving significant accuracy gains and computational savings.

Contribution

The paper proposes TCA, a novel training-free token condensation technique that enhances zero-shot transfer and robustness of vision-language models like CLIP.

Findings

Up to 21.4% performance improvement on cross-dataset benchmarks.

Reduces GFLOPs by 12.2% to 48.9%.

Minimal hyperparameter dependency.

Abstract

Contrastive Language-Image Pretraining (CLIP) excels at learning generalizable image representations but often falls short in zero-shot inference on certain downstream datasets. Test-time adaptation (TTA) mitigates this issue by adjusting components like normalization layers or context prompts, yet it typically requires large batch sizes and extensive augmentations, leading to high computational costs. This raises a key question: Can VLMs' performance drop in specific test cases be mitigated through efficient, training-free approaches? To explore the solution, we investigate token condensation (TC) techniques, originally designed to enhance vision transformer efficiency by refining token usage during inference. We observe that informative tokens improve visual-text alignment in VLMs like CLIP on unseen datasets. However, existing TC methods often fail to maintain in-distribution performance when reducing tokens, prompting us to ask: How can we transform TC into an effective ``free-lunch'' adaptation strategy for VLMs? To address this, we propose Token Condensation as Adaptation (TCA), a training-free adaptation method that takes a step beyond standard TC. Rather than passively discarding tokens, TCA condenses token representation by introducing reservoir-based domain anchor tokens for information-preserving token reduction and logits correction. TCA achieves up to a 21.4% performance improvement over the strongest baseline on cross-dataset benchmark and the CIFAR-100-Corrupted dataset while reducing GFLOPs by 12.2% to 48.9%, with minimal hyperparameter dependency on both CLIP and SigLIP series.