LLM-Derived Priors for Thompson Sampling in Cold-Start Comment Recommendation

In online recommendation systems, new items often suffer from a cold-start problem where insufficient interaction history prevents effective ranking. This paper investigates whether Large Language Models (LLMs) can extract semantic signals from user-generated comments to construct informative Bayesian priors for Thompson sampling, thereby accelerating the learning process in sparse-feedback environments.

The authors propose a segment-aware Thompson sampling framework that incorporates LLM-derived priors. They evaluate two distinct prior designs in a real-world A/B/C test on a webtoon platform:

• Gender Prior: Captures demographic-affinity cues to align comments with specific gender-age segments.
• Content Prior: Identifies title-specific identity cues to highlight comments that best express the unique themes of a work.

These priors are converted into Beta distribution pseudo-counts. The system maintains these posteriors at the demographic-segment level, updating them hourly with real-time impression and click feedback. This hybrid approach allows the system to leverage semantic knowledge for initial cold-start serving while transitioning to data-driven behavioral learning as evidence accumulates.

The study demonstrates that LLM-based priors provide the most significant performance gains in sparse-feedback regimes, particularly after a small amount of interaction evidence has been collected. The researchers found that the two prior designs induce distinct effects across the recommendation funnel: the Gender Prior shows stronger click-oriented alignment, while the Content Prior excels at highlighting thematic identity. Furthermore, the analysis reveals substantial demographic heterogeneity, indicating that the impact of these priors varies significantly across different age and gender groups.

This research provides a practical, scalable mechanism for integrating LLM-based semantic insights into existing bandit-based recommendation pipelines. By using LLMs to initialize priors rather than replacing the bandit policy entirely, the framework maintains the benefits of online adaptation while mitigating the cold-start penalty. The findings highlight the importance of prior design in balancing exploration and exploitation in text-rich recommendation environments.