Hermes: Learning Contextual Reasoning Unlocks Test-Time Scaling

The research shows that the ability of models to effectively distribute computational power across context windows during inference is defined as 'contextual reasoning' and that this skill is learnable. Models trained with the Hermes-Learn framework improve their performance by better leveraging additional computational power.