Xiaomi's A-RSI Boosts Zero-Data Multimodal Self-Improvement via Adaptive Task Refresh
xiaomi-research · hf · 2026-10-30
A-RSI, a paper from Xiaomi Research with CASIA, CUHK, Harvard and others, tackles two core questions in recursive self-improvement: when to introduce fresh training experience and how to keep supervision trustworthy.
- Training dynamics: it monitors reward variance and accuracy in training logs and refreshes a task pool whenever its learning signal declines, without consulting validation or test scores.
- Verifiable supervision: a static library of model-authored programs generates multimodal questions with reference answers; a generation contract plus independent answer auditing keeps the loop independently checkable.
- No external labels: the training loop uses no external examples or human-written answer labels. The authors report that 800 fresh tasks raise average performance across eleven benchmarks from 57.77 to 61.05.
More from Research
- Investor's Last Exam Benchmark: Top AI Models Ace Research but Fail at Investment Judgment — Huangfei Jiang · 2026-10-30
- SpecSlot Boosts Pixel-Space Image Generation: +2.84 FID on ImageNet-512 at Lower Cost — Jiongfan Yang · 2026-10-30
- Apple Trains UI Navigation Agents with Online RL Directly on Live iOS Apps — apple · 2026-10-30
- KAIST's L3G-World Generates Infinite Interactive Worlds in Real Time with Latent 3D Gaussian Memory — kaist-ai · 2026-10-30
- PhysInsert from Beihang and Meitu Achieves Physically Plausible Reference-Based Object Insertion — Yue Gong · 2026-10-30
- AB-Attention Speeds Up Video Diffusion Transformers with Training-Free Accurate Block-Sparse Attention — Bin Xu · 2026-10-30