Tabular Foundation Models Enhance Contextual Bandit Performance.

Contextual bandits are a powerful framework for personalized decision-making, allowing systems to learn optimal actions based on specific user contexts while minimizing exploration costs. However, practical deployment often faces challenges like sparse data, unreliable uncertainty estimates, and the "cold start" problem where little initial data is available. Researchers propose BC-ICL (Bootstrap-conditioned action selection using In-Context Learning) to address these issues. This novel policy transforms pre-trained tabular foundation models, typically used for in-context learning, into randomized policies for online decision-making. In each round, BC-ICL generates a bootstrap resample of past interactions, conditions the frozen foundation model on this resample, scores all available actions, and then selects the action with the highest sampled score. The approach also incorporates an arm-context conditioning architecture, which promotes shared statistical strength across different actions and helps mitigate common failure modes associated with isolated-arm bandits. Empirical evaluations on standard contextual bandit benchmarks demonstrate that BC-ICL delivers strong performance, particularly in early rounds, outperforming established baselines under strict online protocols. This suggests a promising new direction for leveraging large pre-trained models in personalized recommendation and online optimization.