A-RSI: Adaptive Task Refresh for Zero-Data Multimodal Self-Improvement
With 800 fresh tasks, A-RSI raised average performance across eleven benchmarks from 57.77 to 61.05.
The method uses reward variation and training accuracy to decide when a task pool’s learning signal has declined and needs refreshing. Those decisions do not consult validation or test scores. A fixed library of model-authored programs generates multimodal questions and derives their reference answers, which are audited separately. The training loop uses no external training examples or human-written answer labels.
HF Daily Papers' note
The method uses reward variation and training accuracy to decide when a task pool’s learning signal has declined and needs refreshing. Those decisions do not consult validation or test scores. A fixed library of model-authored programs generates multimodal questions and derives their reference answers, which are audited separately. The training loop uses no external training examples or human-written answer labels.
HF Daily Papers' note
score 5