A-RSI Refreshes Stale Tasks, Lifts Multimodal Scores

A-RSI targets a real bottleneck in recursive self-improvement: generating more exercises is easy; knowing when to discard stale ones is harder. Its results are promising, but the evidence supports a practical refresh heuristic more strongly than a breakthrough in adaptive curriculum design.

• –Programmatic answer generation reduces the risk of models training on their own incorrect majority votes.
• –Fresh tasks helped continued GRPO training outperform simply reusing the original pool.
• –The detector’s timing advantage remains unproven: executed refreshes aligned with a fixed three-window schedule.
• –Reported gains partly come from reducing truncated answers, so higher benchmark scores do not necessarily equal deeper reasoning improvements.
• –Eight selection strategies underperformed random task draws, suggesting task diversity may matter more than sophisticated difficulty targeting.