Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..
Buffer layers for Test-Time Adaptation
Authors: Hyeongyu Kim, GeonHui Han, Dosik Hwang
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Through comprehensive experimentation, we demonstrate that our approach not only outperforms traditional methods in mitigating domain shift and enhancing model robustness, but also exhibits strong resilience to forgetting. Furthermore, our Buffer layer is modular and can be seamlessly integrated into nearly all existing TTA frameworks, resulting in consistent performance improvements across various architectures. These findings validate the effectiveness and versatility of the proposed solution in real-world domain adaptation scenarios. |
| Researcher Affiliation | Academia | Hyeongyu Kim1 Geonhui Han1 Dosik Hwang1,2,3 1School of Electrical and Electronic Engineering, Yonsei University 2Department of Radiology and Center for Clinical Imaging Data Science, College of Medicine, Yonsei University 3Artificial Intelligence and Robotics Institute, Korea Institute of Science and Technology {lion4309, hgh6945, EMAIL} |
| Pseudocode | Yes | Algorithm 1 Test-Time Adaptation with Buffer layer |
| Open Source Code | Yes | The code is available at https://github.com/hyeongyu-kim/Buffer_TTA. |
| Open Datasets | Yes | We evaluated our method across several widely used TTA benchmarks, including CIFAR10-C, CIFAR10-W, CIFAR100-C, and Image Net-C... All experiments on CIFAR10-C, CIFAR100-C, and Image Net-C [5] are conducted under Severity 5 settings... CIFAR-10-W [22] is a web-collected dataset... |
| Dataset Splits | Yes | All experiments on CIFAR10-C, CIFAR100-C, and Image Net-C [5] are conducted under Severity 5 settings, following standard protocol to evaluate robustness under the highest level of corruption. ... For the experiment in Section 4.2.6 (Continuously Changing Domains), we use a batch size of 16 during adaptation... |
| Hardware Specification | Yes | All experiments are performed using NVIDIA RTX A6000 GPUs. |
| Software Dependencies | No | The paper mentions software like Adam optimizer, TENT, De Yo, SAR, CMF, and ROID, but does not provide specific version numbers for any of these or for general programming languages or libraries. |
| Experiment Setup | Yes | All experiments including TENT, De Yo, SAR, CMF, and ROID are optimized using the Adam optimizer with a learning rate of 1e-3, β = 0.9, and zero weight decay. EATA shares the same optimizer configuration as TENT (Adam, LR=1e-3, β = 0.9, WD=0.). Additionally, we set the Fisher regularization strength to 1.0 and the confidence margin d to 0.4, following the original implementation. For the source distribution sampling, we use 2000 samples to compute the Fisher information matrix. |