CaRE Scales Continual Learning to 300 Tasks
机器之心 · wechat · 2026-07-13
A research team from HKU proposed a new continual learning method, CaRE (Scalable Continual Learner with efficient Bi-Level Routing Mixture-of-Experts), aiming to scale continual learning to an ultra-long sequence of 300+ non-overlapping tasks.
The core method is BR-MoE: introducing a two-stage routing mechanism within each Transformer block. It first dynamically selects the most relevant task routes based on the entropy of class-aware perceptrons, then selects expert adapters within the corresponding routes for fusion, while retaining a shared expert updated via EMA. The team also constructed a new benchmark, OmniBenchmark-1K, selecting 1,000 classes and about 190,000 images from OmniBenchmark-V2, covering 21 visual domains to evaluate continual learning capabilities across 100 to 300 tasks. Experiments show that CaRE performs robustly across 100/151/200/301 task sequences and maintains a lead on shorter sequence benchmarks like ImageNet-R, ImageNet-A, and ObjectNet.
More from Research
- Partial RoPE study: rotating ~10% of dims matches full RoPE convergence with 10x smaller cache — BlancheMinerva · 2026-10-26
- Spotting reward hacking in benchmarks is far harder than you think, says AI2 researcher — soldni · 2026-10-26
- RCT: ChatGPT temporarily boosts consultants' data science skills, gains vanish without it — EricTopol · 2026-10-26
- FocusVLA routes robot camera views by predicted interaction intent across task stages — iscas-cloud · 2026-10-26
- Transformers' biggest audio overhaul: unified AudioProcessor architecture — LysandreJik · 2026-10-26
- Free 195-page primer: The Little Book of Generative AI Foundations — caglar_ee · 2026-10-26