$ cat wiki/papers/2026/2610.38568-a-rsi.md
A-RSI — refreshing training tasks without using test scores
TL;DR
A-RSI separates when to refresh training experience from how to establish its answers. It uses training dynamics to refresh a multimodal task pool and program-derived reference answers instead of model-majority labels. (source)
Authors & Org
Yu Liu, Youxing Li, Yuedong Tan, Zhiwei Yang, Zihan Su, Hanqi Feng, Shaojie Zhang, Libin Liao, Zihao Wan, Chenxu Wu, Cong Cao, Fangfang Yuan, Yanbing Liu, Ziheng Chen, Pei Fu, Radu Timofte, Zongwei Wu, Nicu Sebe, Barnabas Poczos and Jian Luan. Affiliations are not established in the captured abstract record. First submitted 2026-10-28. (source)
Method
Reward variation and accuracy in training logs trigger refresh when the current pool loses learning signal, without consulting validation or test scores. A static library of model-authored programs generates multimodal questions and derives answers. The authors specify a task-generation contract, automatically verify selected properties of reward and refresh rules, and separately audit generated answers. The training loop uses no external training examples or human-written answer labels. (source)
Results
On the primary eleven-benchmark suite, the authors report the average increasing from 57.77 to 61.05 with 800 fresh tasks. These are reported suite averages, not a general success rate or evidence of unlimited recursive improvement. (source)
Significance
For Agentic Reinforcement Learning, refreshing experience and validating its supervision are different responsibilities. This complements RSIHub: comparing agent evolution under fixed evaluation: a changing training process still needs an evaluation boundary that it does not optimize directly. (source)
Open Questions
The abstract does not name the evaluated backbone, quantify total compute or establish the breadth of the answer audit. “Zero-data” concerns the stated training loop, not the absence of model pretraining or human-designed task contracts. Full-paper details and independent reproduction were not reviewed. (source)