Turing Test Network boosts pseudo-label accuracy by 28% without retraining

A new paper from Brent Griffin and Jason Corso at the University of Michigan presents the Label Imitation Game (LIG), a zero-shot pseudo-label pruning framework that reframes label filtering as an adversarial interrogation. The core innovation is a Turing Test Network (TTN), a task-agnostic “judge” that evaluates candidate pseudo-labels within a dataset-wide context instead of relying on isolated confidence thresholds. This approach targets a key problem in foundation model labeling: hallucinations that slip past standard filters and degrade downstream performance.

The LIG method proves remarkably robust across four diverse datasets, consistently improving label accuracy for three state-of-the-art vision-language models without costly supervision or retraining. The team reports F1-score gains of 28% for the worst-performing baseline categories and 44% when combining with task-specific fine-tuning. Even more striking is the “Category Revival” effect: TTN pruning detoxifies the training signal, enabling downstream models to recover from zero recall on transfer-vulnerable classes. The system also demonstrates unique zero-shot task transfer — a TTN trained solely on image classification datasets can effectively prune complex object detection pseudo-labels, suggesting learned semantic-contextual logic outperforms traditional spatial-geometric verification. The code and pretrained models are publicly available.