Adaptive task refresh improves zero-data multimodal model training by detecting learning saturation

Recursive self-improvement in machine learning involves models generating their own training experience to improve iteratively. A key challenge is not only deciding what new tasks to generate but also when to introduce fresh tasks, as continuing to train on a saturated task pool wastes updates, while refreshing too often can consume unnecessary new data.

Existing methods often focus on selecting or generating tasks based on difficulty or model competence but pay less attention to detecting when the current training data no longer provides useful learning signals. Additionally, many approaches rely on self-consistency or majority voting to generate training targets, which can propagate the model's own uncertain predictions as supervision, risking error reinforcement.

Separating the timing of task refresh from the correctness of task answers is important. The model should use training dynamics to decide when to refresh, while the new tasks should have independently verifiable answers to avoid supervision based on uncertain model outputs. This separation is central to the approach in this paper.