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..
TRiCo: Triadic Game-Theoretic Co-Training for Robust Semi-Supervised Learning
Authors: Hongyang He, Xinyuan Song, Yangfan He, Zeyu Zhang, Yanshu Li, Haochen You, Lifan Sun, Wenqiao Zhang
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments on CIFAR10, SVHN, STL-10, and Image Net demonstrate that TRi Co consistently achieves state-of-the-art performance in low-label regimes, while remaining architectureagnostic and compatible with frozen vision backbones. |
| Researcher Affiliation | Collaboration | Hongyang He1 , Xinyuan Song2 , Yangfan He3, Zeyu Zhang4, Yanshu Li5, Haochen You6, Lifan Sun7, Wenqiao Zhang81University of Warwick 2Emory University 3University of Minnesota Twin Cities 4ANU 5Brown University 6Columbia University 7UCSD 8Zhejiang University Manifolda.AiEqual contribution. Work was done while the authors were with the Manifolda.Ai Research team. Corresponding author: EMAIL. |
| Pseudocode | Yes | Algorithm 1 TRi Co: Triadic Game-Theoretic Co-Training |
| Open Source Code | No | The paper does not provide an explicit statement or link for open-source code release in its main body or appendices. |
| Open Datasets | Yes | All models are trained on four benchmark datasets: CIFAR-10 [28], SVHN [33], STL-10 [14], and Image Net [18]. |
| Dataset Splits | Yes | For CIFAR-10 and SVHN, we follow standard semi-supervised settings using 4,000 labeled examples (10%) and test on the full test set. For STL-10, we use all labeled and 100k unlabeled samples. On Image Net, we evaluate two settings with 25%, 10% and 1% labeled subsets respectively, following protocols in Sohn et al. [39]. Across all datasets, we use a fixed labeled validation split (10% of labeled data) to compute the meta-gradient for teacher updates. |
| Hardware Specification | Yes | We implement TRi Co in Py Torch using 4 NVIDIA A100 GPUs. [...] Runs on a single RTX 3090/A6000 GPU |
| Software Dependencies | No | The paper mentions software like Py Torch, but does not specify version numbers for any key software components. |
| Experiment Setup | Yes | Student models are trained using SGD with momentum 0.9 and batch size 64, with a cosine learning rate decay starting from 0.03. The teacher parameters (τMI, λu, λadv) are updated via meta-gradient descent using a labeled validation batch, with initial values (0.05, 0.5, 0.5) and learning rate 0.01. Each experiment is run for 512 epochs, and we report mean accuracy over 3 random seeds. |