Skill Entropy reward boosts multi-skill reasoning

Quantifying how hard a model is to switch between reasoning skills produces dramatic accuracy gains...