A-RSI: Adaptive task refresh for zero-data multimodal self-improvement
In brief
- A-RSI automatically refreshes the task pool when learning signals decline, without external validation.
- The system generates multimodal questions and reference answers via static model-authored programs.
- On 11 benchmarks, average performance rises from 57.77 to 61.05 with 800 new tasks.
Researchers introduce A-RSI, a zero-data multimodal self-improvement method. The system monitors reward and accuracy variations in training logs to decide when to refresh the task pool, without relying on validation or test scores. A static library of model-authored programs generates multimodal questions and their reference answers, which are automatically verified against specified properties. The training loop uses no external examples or human-written labels.
On an eleven-benchmark suite, A-RSI raises the average performance from 57.77 to 61.05 with 800 fresh tasks, without human supervision or external data.
This summary gives you the essentials. The full article, its figures and quotes are on Hugging Face Daily Papers.
Read the original articleAI-generated summary of the source, checked by rules. How we write · Report an error