An image-adaptation method called ReNC was reported as the top performer in benchmark comparisons where streams mixed altered examples from known categories with images from outside those categories, known as out-of-distribution (OOD) data. A rank-based Friedman test rejected the assumption that the methods performed equally, and a follow-up Nemenyi comparison favored ReNC over the baselines.
How ReNC adapts on the fly
ReNC is a source-free test-time adaptation method: it updates a source-trained classifier as target images arrive, without using labeled target examples. It uses what the authors call neural-collapse geometry—a pattern in a model's internal feature space—as a structural guide. ReNC first filters likely OOD samples and then stochastically updates class prototypes, the model's representative vectors for known categories, to approximate that geometry in the target domain.
Researchers judged the trade-off with three measures: ACCI, accuracy on in-distribution (ID) images; ACCO, accuracy on OOD images; and ACCH, a harmonic mean, meaning a combined score pulled down when either accuracy is low.
A test built around mixed image streams
The evaluation used five separate ID benchmarks rather than one pooled sample: CIFAR10-C and CIFAR100-C each had 10,000 images, ImageNet-C had 50,000, ImageNet-R had 30,000 and VisDA-C had 55,388. OOD sources included Noise, MNIST, SVHN, Tiny-ImageNet, CIFAR100-C and CIFAR10-C, and the reported OOD-to-ID ratio was 1. The paper does not report a pooled stream total or OOD sample counts.
Scores varied by benchmark
On CIFAR10-C, ReNC's ACCH score ranged from 72.58 when the OOD source was CIFAR100-C to 89.47 with MNIST, with the other tested sources between those values. On CIFAR100-C, scores ranged from 48.22 with Tiny-ImageNet to 65.16 with Noise.
On ImageNet-C, ReNC's reported scores were between 52.35 and 53.94 across Noise, MNIST and SVHN. On the style-transfer benchmarks, scores ranged from 54.65 to 56.77 on ImageNet-R and from 64.16 to 71.67 on VisDA-C.
In a continual CIFAR10-C test, ReNC's five ACCH values averaged 82.26; under a ViT backbone, its five values ranged from 82.72 to 94.88. The paper reports the highest continual-benchmark average and says ReNC remained highest in the ViT comparison. In an ImageNet-R vision-language-model extension, ReNC was ranked first and C-TPT+ second, with scores ranging from 54.65 to 56.77 for ReNC and from 40.38 to 45.39 for C-TPT+.
What the benchmark leaves open
In an ablation, the full loss-and-prototype-update configuration had the strongest reported ACCH scores in the cited CIFAR10-C and CIFAR100-C rows. The authors treat the method's neural-collapse-related representation measures as a possible explanation; in one target-domain analysis, ReNC had the lowest reported NC1 and NC3 on CIFAR10-C with Noise and MNIST OOD samples while remaining competitive elsewhere. Those measures track representation structure, however, and do not by themselves show that the structure caused the accuracy results.
The evidence is limited to a version-1 arXiv preprint and benchmark evaluation, not human or real-world deployment. The method assumes a sufficiently trained source model with approximately neural-collapsed classifier weights, and the reported results include no confidence intervals or run-level variability estimates.
Several non-OWTTA baselines were adapted using the paper's own ID/OOD partitioning protocol, while OWT3 was modified for source-free evaluation, so comparisons depend on that setup. ReNC also assigns OOD images to one additional category rather than distinguishing multiple unseen classes. Whether it holds when the source model lacks the assumed geometry, class coverage is highly imbalanced or the results are repeated on independent datasets remains open.
Paper data and sources
Original title: Reliable Neural Collapse Approximation for Open-World Test-Time Adaptation
Authors: Jia-Qi Lin, Yuangang Pan, Chang-Dong Wang et al.
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-20
DOI: Not available
Original paper · Full text