Reproduce “A-RSI: Adaptive Task Refresh for Zero-Data Multimodal Self-Improvement” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Yu Liu, Youxing Li, Yuedong Tan, Zhiwei Yang, Zihan Su, Hanqi Feng, Shaojie Zhang, Libin Liao and 12 more
A-RSI uses training logs to replace a task pool with unused program-generated tasks while preserving learned weights. Its programs independently compute answers, while its trigger tracks reward-signal decline. In one Qwen3-VL-8B-Instruct run, Macro11 accuracy rose from 57.77 to 61.05 after eight windows with 800 fresh tasks. This supports the combined refresh policy, but does not establish optimal timing or mastery.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.