MULTILINGUAL (SPANISH) Lastly, one interesting pattern we noticed is that every model seemed to perform worse at Instruction Following (Main Request fulfillment and Constraint fulfillment) in Spanish versus English. This implies there's still meaningful headroom for models to