A-RSI refreshes tasks for zero-data multimodal self-improvement
Original title: A-RSI: Adaptive Task Refresh for Zero-Data Multimodal Self-Improvement
On an eleven-benchmark suite the authors report the average rising from 57.77 to 61.05 using 800 fresh tasks.
A-RSI is a recursive self-improvement method that decides when current training experience has stopped helping, rather than only what to generate next. It watches reward variation and accuracy in training logs and refreshes a task pool when the learning signal declines, without using validation or test scores. A static library of model-authored programs creates multimodal questions and independently checkable reference answers, and the loop uses no external training examples or human-written labels. On the primary eleven-benchmark suite, the authors report the performance average rising from 57.77 to 61.05 with 800 fresh tasks.
- Refresh timing uses training-log reward variation and accuracy, not held-out scores.
- Reference answers come from model-authored programs and are audited separately.
- The training loop uses no external examples or human-written answer labels.
- Reported average moves from 57.77 to 61.05 with 800 fresh tasks.
From the original article
1 min read at huggingface.co
Recursive self-improvement requires models not only to acquire useful training experience, but also to decide when their current experience has ceased to be useful. Existing methods often focus on what experience to generate next, while paying less attention to when fresh experience should be introduced.
By Yu Liu, Youxing Li, Yuedong Tan · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.