MoDE Ensemble Treats Late LLM Layers as Tunable Predictors

Late transformer layers already carry largely-formed predictions, and MoDE exploits this by routing final logits through a learned ensemble of those layers. It stacks on top of existing methods like LoRA with minimal parameter overhead and shows modest reasoning gains in experiments.