Calibration-gated LLM pseudo-observations reduce cumulative regret by 19% versus pure LinUCB on a 5-arm news recommendation task when using task-specific prompts, but generic prompts increase regret on both tested environments.
arXiv preprint arXiv:2406.02611
2 Pith papers cite this work. Polarity classification is still indexing.
2
Pith papers citing it
verdicts
UNVERDICTED 2representative citing papers
RecPIE jointly optimizes recommendation predictions and LLM-generated natural-language explanations via alternating training and reinforcement learning, yielding 3-4% accuracy gains and higher human preference on Google Maps POI data.
citing papers explorer
-
Calibration-Gated LLM Pseudo-Observations for Online Contextual Bandits
Calibration-gated LLM pseudo-observations reduce cumulative regret by 19% versus pure LinUCB on a 5-arm news recommendation task when using task-specific prompts, but generic prompts increase regret on both tested environments.
-
Can Explanations Improve Recommendations? Evidence from Prediction-Informed Explanations
RecPIE jointly optimizes recommendation predictions and LLM-generated natural-language explanations via alternating training and reinforcement learning, yielding 3-4% accuracy gains and higher human preference on Google Maps POI data.