Research map / Label-efficient and robust learning / Semi-supervised learning

TRiCo: Triadic Game-Theoretic Co-Training for Robust Semi-Supervised Learning

NeurIPS 2025 · Label-efficient and robust learning › Semi-supervised learning · approach: Confidence-based pseudo-label selection

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes TRiCo, a triadic game-theoretic co-training SSL framework (two students on frozen views, a meta-learned teacher, and an embedding-space adversarial generator) that filters pseudo-labels by mutual information and reports state-of-the-art SSL accuracy on CIFAR-10, SVHN, STL-10 and ImageNet.

Problem

Semi-supervised learning under low label budgets suffers from brittle confidence-based pseudo-label selection, static symmetric view interactions in co-training, and little mechanism for surfacing hard, boundary-near examples.

Approach

Combine two student classifiers on frozen complementary ViT embeddings with a meta-learned teacher that sets an MI threshold and loss weights via validation-loss meta-gradients, plus FGSM-style embedding perturbations, and frame the interaction as a Stackelberg/Nash game.

Key results

• TRiCo reports 96.3% on CIFAR-10 (4k labels) and 94.2% on SVHN (1k labels), above FixMatch, FlexMatch and Meta Pseudo Label in its table.
• On ImageNet, TRiCo reports 81.2% at 1%, 85.9% at 10%, and 88.3% at 25% labels; it is below REACT (81.6%) at 1%.
• Ablations on CIFAR-10 report MI filtering at 95.9% accuracy versus 95.8% for the best confidence threshold (0.75), a small gap.

Measured effects reported in the paper

Surprising findings

• The paper's own ablation shows MI filtering at 95.9% accuracy vs 95.8% for the best confidence threshold (0.75), so the advantage over tuned confidence thresholds is marginal.
• The reported 1% ImageNet result (81.2%) is below REACT (81.6%), which the paper's text acknowledges.

Limitations

• The meta-gradient for the teacher is a first-order one-step unroll; the theoretical Nash existence proof relies on loose assumptions (e.g., continuity of 0-1 loss surrogates) and does not show convergence of training.
• Several reported numbers and settings are internally inconsistent (e.g., the teacher was said to use an MI threshold but ablations list lambda_MI; 512 epochs and claimed PGD robustness lack detail), and code or data release is not stated.

Experimental setup

Datasets: CIFAR-10 SVHN STL-10 ImageNet CIFAR-100 CUB-200-2011 CIFAR-10-C CIFAR-100-LT

Benchmarks: CIFAR-10 SVHN STL-10 ImageNet CIFAR-100 CUB-200-2011 CIFAR-10-C CIFAR-100-LT

Metrics: top-1 accuracy test error mCE PGD robustness accuracy

Baselines: FixMatch UDA FlexMatch Meta Pseudo Label Semi-ViT REACT MCT SimCLRv2+KD FreeMatch SemiFormer

Builds on: FixMatch Meta Co-Training Meta Pseudo Label co-training

Closest papers by meaning

• AT-Field: Rethinking the Games in Adversarial Training (AAAI 2026 · similarity 0.85)
• Revisiting Semi-Supervised Learning in the Era of Foundation Models (NeurIPS 2025 · similarity 0.84)
• Toward Improving Robustness and Accuracy in Unsupervised Domain Adaptation (AAAI 2025 · similarity 0.84)
• Rethinking Confidence Scores and Thresholds in Pseudolabeling-based SSL (ICML 2025 · similarity 0.83)
• Contrastive Credibility Propagation for Reliable Semi-supervised Learning (AAAI 2024 · similarity 0.83)
• Seirênes: Adversarial Self-Play with Evolving Distractions for LLM Reasoning (NeurIPS 2026 · similarity 0.83)