Meta FAIR Proposes a Routed Continual-Learning Architecture
Meta FAIR proposed a Routed Continual Learning (Routed CL) architecture. [[fact:fact-8vsi8twm2]]The architecture groups historical trajectories into clusters and maintains a separate memory for each cluster. [[fact:fact-c47xs2ap6]]Compared with Global ICL, Routed CL’s normalized-reward improvement was equivalent to 15% to 30% of the maximum achievable normalized reward, while estimated API costs fell by 65% to 72%. [[fact:fact-zsx564sls]]Across four models, Routed CL had higher average normalized rewards than Stateless ICL, Global ICL, and Routed ICL. [[fact:fact-6naihi3ls]]The current experiments process tasks sequentially, with one language model acting as the executor in each rollout and taking on all model-mediated roles in the memory system. [[fact:fact-amvll95pw]]