Sakshi Arya

Contextual bandits don't actually need reward estimation—exploiting the elicitation geometry of shared monotone links lets algorithms target the decision boundary directly to secure logarithmic regret under severe functional misspecification.