Calibration-Gated LLM Pseudo-Observations for Online Contextual Bandits

TL;DR

This paper introduces a method to improve online contextual bandit algorithms by using large language models to generate pseudo-observations, with a calibration-based weighting scheme to enhance early decision-making.

Contribution

It proposes a novel augmentation of Disjoint LinUCB with LLM pseudo-observations and a calibration-gated decay schedule to mitigate cold-start regret.

Findings

LLM pseudo-observations reduce regret by 19% on MIND with task-specific prompts.

Prompt design significantly impacts performance, more than decay schedule or gating parameters.

Calibration gating's effectiveness varies with prediction error levels, affecting bias-variance trade-offs.

Abstract

Contextual bandit algorithms suffer from high regret during cold-start, when the learner has insufficient data to distinguish good arms from bad. We propose augmenting Disjoint LinUCB with LLM pseudo-observations: after each round, a large language model predicts counterfactual rewards for the unplayed arms, and these predictions are injected into the learner as weighted pseudo-observations. The injection weight is controlled by a calibration-gated decay schedule that tracks the LLM's prediction accuracy on played arms via an exponential moving average; high calibration error suppresses the LLM's influence, while accurate predictions receive higher weight during the critical early rounds. We evaluate on two contextual bandit environments - UCI Mushroom (2-arm, asymmetric rewards) and MIND-small (5-arm news recommendation) - and find that when equipped with a task-specific prompt, LLM pseudo-observations reduce cumulative regret by 19% on MIND relative to pure LinUCB. However, generic counterfactual prompt framing increases regret on both environments, demonstrating that prompt design is the dominant factor, more important than the choice of decay schedule or calibration gating parameters. We analyze the failure modes of calibration gating on domains with small prediction errors and provide a theoretical motivation for the bias-variance trade-off governing pseudo-observation weight.