Search Results for "labels"

TTRL is an open-source framework for test-time reinforcement learning in large language models, with a particular focus on reasoning tasks where ground-truth labels are not available during inference. The project addresses the problem of how to generate useful reward signals from unlabeled test-time data, and its central insight is that common test-time scaling practices such as majority voting can be repurposed into reward estimates for online reinforcement learning. This makes the framework especially interesting for scenarios where models must keep adapting during evaluation or deployment instead of relying only on fixed pretraining and static fine-tuning. ...