# A-RSI: Adaptive Task Refresh for Zero-Data Multimodal Self-Improvement

**Authors:** 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, Jian Luan  
**Lab:** Xiaomi Research · **Published:** 27 Oct 2026 · **Upvotes:** 163 · **Level:** Advanced  
**Links:** [arXiv](https://arxiv.org/abs/2610.38568) · [PDF](https://arxiv.org/pdf/2610.38568) · [This page](https://omen-bit.github.io/ai-paper-daily/paper/2610.38568.html)

## Curator's notes

**TL;DR:** A-RSI adaptively adds new self-generated tasks when learning signals drop, improving zero-data multimodal self-improvement.

- **What's new:** It introduces a reward‑variation based refresh mechanism and separates answer verification from task generation, unlike prior methods that rely on self‑consistency.
- **Key result:** On eleven benchmarks, A-RSI raises the average performance from 57.77 to 61.05 using 800 fresh tasks.
- **Why it matters:** Builders can get better zero‑data models without human labels by automatically knowing when to create new training tasks.

## Key numbers

- **800**: A-RSI raises the performance average from 57.77 to 61.05 with 800 fresh tasks

## Abstract

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. Many also construct training targets through self-consistency or majority voting, making supervision depend on the model's own uncertain predictions. We address these two questions separately: training dynamics determine when to acquire new experience, while the answers to new tasks remain independently checkable. A-RSI monitors reward variation and accuracy in training logs to refresh a task pool when its learning signal declines, without consulting validation or test scores. A static library of model-authored programs produces multimodal questions and derives their reference answers; we specify the task-generation contract, automatically verify chosen properties of the reward and refresh rules, and separately audit the generated answers. The training loop uses no external training examples or human-written answer labels. On the primary eleven-benchmark suite, A-RSI raises the performance average from 57.77 to 61.05 with 800 fresh tasks.

## What you could build

1. Generate a self‑improving chatbot that adds new conversation prompts when reward variance falls.
2. Create a multimodal QA system that auto‑refreshes image‑question pairs as accuracy plateaus.
3. Develop a code‑generation assistant that expands its task pool based on training loss trends.

## Quiz yourself

What metric does A-RSI monitor to decide when to refresh the task pool?

**Answer:** A-RSI monitors reward variation and accuracy in the training logs to determine when the learning signal has declined.

---

Notes written by AI from the paper, numbers copied from the abstract. Check the paper before you rely on them. From AI Papers Daily.