LLM priors warm-start Thompson sampling, boost comment clicks

Cold-start remains a thorny problem for multi-armed bandit algorithms like Thompson sampling: newly introduced items, or arms, have little to no interaction history, so exploration is inefficient. This arXiv paper tackles that problem for user-generated comments, where the text itself carries signals about appeal before any engagement. The authors use large language models to extract semantic cues from comment content and convert them into informative Bayesian priors, effectively warm-starting Thompson sampling under sparse early-stage feedback. To handle demographic differences, they maintain and update separate posteriors for each gender-age segment.

The team ran a real-world online A/B/C test comparing a uniform prior against two LLM-derived designs: a Gender Prior, capturing demographic-affinity signals, and a Content Prior, capturing title-specific identity cues. Results show LLM priors are most valuable in low-feedback conditions, with the biggest gains emerging once a small amount of interaction evidence accumulates. Interestingly, click-oriented alignment was strongest for the Gender Prior, while treatment effects varied substantially across demographic segments. These findings suggest that LLM-derived priors can serve as a practical warm-start mechanism for text-rich bandit recommendation, but deployment involves trade-offs—prior design influences funnel-level metrics differently, and demographic heterogeneity must be accounted for.