An arXiv preprint reports that PuRF, a method for adapting vision-language models to multi-label image recognition, scored higher than comparison methods in the benchmark settings it tested. On ViT-B/32, PuRF showed a reported average mAP improvement of 8.56% over ML-TTA and 4.05% over ReTA across the datasets. On COCO2014, the reported gain over ReTA was 6.09%.
The result comes from a fully unlabeled online setting. Each incoming test sample is handled without labels, while the method adapts as the stream continues. PuRF is described as a purification-driven, cache-based method, meaning its design centers on selecting and updating stored model evidence as new samples appear.
How PuRF adapts as images arrive
The method combines Region Purification with Cache Purification's episodic purification and temporal refreshing. Region Purification is used to identify reliable regions in an image, while the cache operations govern what is retained and refreshed. In ordinary terms, PuRF is designed to make the visual cues entering its cache more selective and to keep that cache current as the image stream changes.
For the reported implementation, the method generated 50 regions and used 63 augmented visual views. Its cache size was set to L = 3, alongside three adjacent text embeddings. The reported settings also included a temporal decay constant of δ = 1000, global filtering of κg = 0.1, and coefficients λ1 = 0.2 and λ2 = 0.5.
Tests across datasets and model choices
The evaluation covered five multi-label datasets: VOC2007, VOC2012, COCO2014, COCO2017 and NUS-WIDE. The reported test sets contained 4.9K images for VOC2007, 5.8K for VOC2012, 40K for COCO2014, 5K for COCO2017 and 83K for NUS-WIDE. The paper lists the test-set size separately for each benchmark.
PuRF was evaluated with four listed vision-language backbones: RN50, RN101, ViT-B/16 and ViT-B/32, with mAP as the reported metric. One reported prompt setup used MaPLe initialization. In that comparison, PuRF reached 73.79% average mAP, compared with 70.88% for ReTA and 70.38% for ML-TTA.
What the reported comparisons showed
The reported advantage also appeared in a comparison using diverse VLM backbones. In that analysis, PuRF's average mAP was 3.56% higher than DPE's and 3.69% higher than ReTA's. The paper presents these figures as average improvements across the backbone comparison, rather than as a result tied to only one model.
One reported efficiency check used a single NVIDIA 3090 GPU and the ViT-B/16 backbone. Under that setup, PuRF reached 5.74 frames per second, used 1.04 GB of memory and recorded 71.48% mAP, alongside an 8.87% reported gain. These figures put speed, memory and recognition score in the same reported test configuration.
A component analysis reported 88.18% mAP on VOC2007 and 65.35% on COCO2017 when Region Purification, Episodic Purification and Temporal Refreshing were used together. Those figures describe the full listed combination of the method's named components.
The scope of the evidence
The findings remain tied to the paper's reported computational settings. The datasets are the five named benchmarks, the model comparisons use the listed backbones, and the implementation uses particular region, augmentation, cache and filtering choices. The reported gains therefore show how PuRF compared in these experiments; they do not, on their own, establish performance on other data streams, distribution shifts or hardware.
The front matter identifies the work as an arXiv preprint, version 1, dated 26 Aug 2026. The study's evidence is a computational comparison of model performance on the named image benchmarks, so the reported percentages should not be read as results beyond those tests.
The work was supported by the National Natural Science Foundation of China and the Beijing Natural Science Foundation.
Paper data and sources
Original title: Towards Purified Multi-Label Test-Time Adaptation of Vision-Language Models
Authors: Yiwen Liang, Hui Chen, Yizhe Xiong et al.
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-26
DOI: Not available
Original paper · Full text