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..
Lifelong Test-Time Adaptation via Online Learning in Tracked Low-Dimensional Subspace
Authors: Dexin Duan, Rui Xu, Peilin Liu, Fei Wen
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments demonstrate that LCo TTA effectively overcomes degeneration and significantly outperforms existing methods in long-term continual adaptation scenarios. We conduct experiments on the Image Net C dataset [53], which consists of 15 corruptions each with 5 severity levels. |
| Researcher Affiliation | Academia | Dexin Duan, Rui Xu, Peilin Liu, and Fei Wen School of Integrated Circuits, School of Information Science and Electrical Engineering Shanghai Jiao Tong University, Shanghai, China, 200240 EMAIL |
| Pseudocode | No | The paper describes its methodology through detailed textual explanations and mathematical formulations, but it does not include any explicitly labeled pseudocode or algorithm blocks. |
| Open Source Code | Yes | Code is available at https://github.com/Thunder David/LCo TTA. |
| Open Datasets | Yes | We conduct experiments on the Image Net C dataset [53], which consists of 15 corruptions each with 5 severity levels. Results on CIFAR100C and semantic segmentation are provided in Appendix K and L. We conduct online continual test-time adaptation experiments using the CARLA simulator [59] across three domain-shift scenarios with varying weather and visual conditions: day-to-night, clean-to-fog, and clean-to-rain. We further evaluate our method on segmentation tasks using the real-world Cityscapes [60] dataset under corrupted target domains for a more intuitive demonstration of its effectiveness. |
| Dataset Splits | Yes | We evaluate our method under a challenging long-term continual TTA setting, where the model adapts continuously over 50 cycles of 15 corruption types (severity=5), a total of 37.5 million test samples. The model performs unsupervised continual adaptation without any external intervention from the very first beginning, such as domain-specific information, model resetting, or warm-up. Each cycle contains 15 corruptions with 50000 samples for each corruption, resulting in a total of 3.75 107 samples used in the 50 cycles. |
| Hardware Specification | Yes | We conduct the main experiments of 50 cycles TTA in Section 6.1 on a Linux server equipped with 8 NVIDIA V100 GPUs with 32GB memory each, and an Intel(R) Xeon(R) Platinum 8280 CPU @ 2.70GHz. All other experiments in Section 6.2 are performed on a PC platform equipped with a single Nvidia RTX 3090 GPU with 24GB memory, including the efficiency analysis in Table 7. |
| Software Dependencies | No | The paper does not explicitly list specific software dependencies with version numbers (e.g., programming language versions, library versions, or framework versions) needed to replicate the experiments. |
| Experiment Setup | Yes | We evaluate our method under a challenging long-term continual TTA setting, where the model adapts continuously over 50 cycles of 15 corruption types (severity=5), a total of 37.5 million test samples. The model performs unsupervised continual adaptation without any external intervention from the very first beginning, such as domain-specific information, model resetting, or warm-up. Effect of subspace dimension. Figure 5 shows the performance of our method with different dimensions of the subspace, r {10, 25, 50, 100}, in continual adaptation on Imagenet C over 50 cycles. Table 5 presents a comparison between Tent and our method across different learning rates (LR) in continual adaptation on Image Net-C over one cycle with Res Net-50. |