Towards Looped Models Done Right, Part I: Topology, Input Injection, Recurrent-State Design

Benchmark Model Rank Results
multi-task-language-understanding-on-mmluHuginn MoE (8.0B resident, 0.8B active)#37Average (%): 66.81
mathematical-reasoning-on-math-500Huginn MoE (8.0B resident, 0.8B active)#123Accuracy: 53.4
arithmetic-reasoning-on-gsm8kHuginn MoE (8.0B resident, 0.8B active)#135Accuracy: 83.62
sentence-completion-on-hellaswagHuginn MoE (8.0B resident, 0.8B active)#165Accuracy: 55.1