Contextual Bandit
A class of sequential decision-making frameworks in reinforcement learning in which an agent selects actions based on contextual features and receives immediate rewards, without involving long-term state transitions or step-by-step credit assignment.
Timeline (last 90 days)
Contextual bandit is a class of sequential decision-making frameworks in reinforcement learning: the agent observes contextual features each round, selects an action and receives only an immediate reward, without involving the step-by-step credit assignment problem and without requiring temporal backup estimation of a value function.
Cord models "given the prompt and the current partial response, selecting a residual coefficient to modulate the revealing order" as a contextual bandit problem, where the action space consists of low-dimensional residual coefficients rather than specific tokens, and the reward is taken from a sequence-level metric after the entire sequence has been generated.
All Facts (2)
Contextual bandit is a class of sequential decision-making frameworks in reinforcement learning: the agent observes contextual features each round, selects an action and receives only an immediate reward, without involving the step-by-step credit assignment problem and without requiring temporal backup estimation of a value function.
50%UnverifiedCord models "given the prompt and the current partial response, selecting a residual coefficient to modulate the revealing order" as a contextual bandit problem, where the action space consists of low-dimensional residual coefficients rather than specific tokens, and the reward is taken from a sequence-level metric after the entire sequence has been generated.
50%