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..
Minimal Semantic Sufficiency Meets Unsupervised Domain Generalization
Authors: Tan Pan, Kaiyu Guo, Dongli Xu, Zhaorui Tan, Chen Jiang, Deshu Chen, Xin Guo, Brian Lovell, LIMEI HAN, Yuan Cheng, Mahsa Baktashmotlagh
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Empirically, MS-UDG sets a new state-of-the-art on popular unsupervised domain-generalization benchmarks, consistently outperforming existing SSL and UDG methods, without category or domain labels during representation learning. ... 5 Experiments |
| Researcher Affiliation | Academia | 1 AI3, Fudan University 2 Shanghai Academy of Artificial Intelligence for Science 3 The University of Queensland EMAIL, EMAIL, EMAIL |
| Pseudocode | No | Section 4.1 is titled 'Algorithm' but describes the method in prose and refers to Figure 1 for a pipeline illustration, not a structured pseudocode block. |
| Open Source Code | No | Question: Does the paper provide open access to the data and code, with sufficient instructions to faithfully reproduce the main experimental results, as described in supplemental material? Answer: [No] Justification: We will open the dataset splits and code upon publication. |
| Open Datasets | Yes | Datasets. Following previous UDG work [56], PACS [23] and Domain Net [35] are evaluated for benchmarking UDG methods. Meanwhile, we also evaluate our methods on Office Home, Office31, and VLCS whose results and details can be found in the Appendix. |
| Dataset Splits | Yes | For Domain Net, the six domains are split into two sets: 1. Clipart, Infograph, and Quickdraw; 2. Painting, Real, and Sketch as previous work did. 20 classes are selected as both unlabeled and labeled data. We use one set for seen domains and the other one for unseen domains. For PACS, three domains are selected for training, and the remaining domain is used for evaluation. ... we adopt linear probing for 1% and 5% label fractions and fine-tuning for 10% and 100% label fractions in downstream tasks. ... Those benchmarks are all adopted leave-one-domain-out validation. |
| Hardware Specification | Yes | All experiments were conducted on a single Nvidia A100 80GB GPU, an 8-core CPU, and 250GB of memory. |
| Software Dependencies | No | The paper mentions 'AdamW' as an optimizer but does not specify any software libraries or frameworks with version numbers (e.g., Python, PyTorch, TensorFlow, CUDA versions). |
| Experiment Setup | Yes | Table 4: Experimental Settings. Lp represents linear probing, and ft represents fine-tuning. label fraction strategy backbone epoch batch_size weight_decay learning_rate optimizer betas Pretraining (Domain Net and PACS) ViT-S/16 80 192 0.05 1e-4 AdamW (0.9,0.95) Finetuning (Domain Net) 1% lp ViT-S/16 50 32 0.05 5e-4 AdamW (0.9,0.95) 5% lp ViT-S/16 50 128 0.05 5e-4 AdamW (0.9,0.95) 10% ft ViT-S/16 50 128 0.05 5e-4 AdamW (0.9,0.95) 100% ft ViT-S/16 50 128 0.05 5e-4 AdamW (0.9,0.95) Finetuning (PACS) 1% lp ViT-S/16 50 16 0.05 5e-4 AdamW (0.9,0.95) 5% lp ViT-S/16 50 64 0.05 5e-4 AdamW (0.9,0.95) 10% ft ViT-S/16 50 64 0.05 5e-5 AdamW (0.9,0.95) 100% ft ViT-S/16 50 64 0.05 5e-5 AdamW (0.9,0.95) |