LLM-Derived Priors for Thompson Sampling in Cold-Start Comment Recommendation
Eugene Lee, Oseong Choi, Byungsoo Kang, Taeyeong Jang
cs.IR, cs.LG
2026-08-04
On Webtoon's comment recommendation, new comments lack click history and the Thompson-sampling bandit cannot cold-start. NAVER uses GPT-4.1 to extract semantic signals from comment text into Beta priors that warm-start the bandit, updating posteriors per gender-age segment. An online A/B/C test shows CTR up to 9.5% higher in sparse-feedback buckets, though aggregate numbers stay flat and exposure concentrates.
The setting is comment recommendation on NAVER Webtoon, a Korean webtoon platform with roughly 4,000 active titles. On the discovery surface, users decide fast from thumbnails, titles, and limited metadata, which only partly convey what a work is like. So the platform recommends reader-written comments that make a title's appeal tangible through tone, character dynamics, and emotional atmosphere.
The policy is Thompson sampling, a multi-armed bandit: each comment is an arm, a click is the reward, and the bandit learns as it serves. The trouble is that this is a newly launched component, so every new comment starts with near-zero click history and an essentially empty posterior, meaning the bandit chooses blindly. That is cold start. One more layer: posteriors are maintained per gender×age segment (2 genders × 4 age bins), because tastes differ sharply across demographic groups.
The idea: before behavioral feedback arrives, use an LLM to extract semantic signals from comment text and turn them into a Bayesian prior that warm-starts the bandit; once real clicks accumulate, the posterior takes over. The LLM (GPT-4.1) runs only offline (re-scored daily for new comments), so online serving makes zero LLM calls and latency is unaffected.
The prior is additive: a shared base score (the comment's universal "hook" strength) plus an adjustment. Two adjustments are tested: the Gender Prior (gender-affinity cue, range −0.35 to 0.40, scaling 2.0) and the Content Prior (title-specific identity cue, range 0 to 1, scaling 1.0). The base prompt asks for a CTR-like score (most comments below 0.5); gender and content prompts estimate adjustments. All prompts output CTR-like quantities used directly as prior means, skipping a calibration step.
A prior mean μ becomes Beta pseudo-counts: α₀=1+round(κμ), β₀=1+round(κ(1−μ)), with prior strength κ=40. This 40 is not a tuned optimum but an experimental value chosen so the prior's effect is observable under sparse traffic; the median comment gets 12 impressions over 7 days and the 75th percentile about 40, so κ=40 keeps the prior influential for most comments while letting behavioral evidence match it near the upper quartile. Posteriors update hourly over a 7-day sliding window: α=α₀+k (clicks), β=β₀+n−k (non-click impressions). At serving time, one posterior sample is drawn per candidate and the top 10 are shown.
A real online A/B/C test, roughly 595K users per variant, over a fixed four-week window (Mar 7 to Apr 3, 2026). A = uniform prior (control), B = Gender Prior, C = Content Prior.
At the aggregate level the results are flat or even negative: B lifts CTR +1.48% and CVR +1.23%, neither significant; C significantly lowers CTR by 5.68% (p<0.001) with a directionally positive but non-significant CVR (+1.37%). The two priors act at different funnel stages: the Gender Prior leans toward immediate clicks, the Content Prior toward downstream conversion.
The value shows up only after slicing. Grouping by a comment's cumulative impressions, sparse-feedback buckets gain the most:
| Impression bucket | Gender Prior CTR lift | Content Prior CTR lift |
| 0–9 | Directionally positive, n.s. | Directionally positive, n.s. |
| 10–49 | +9.51% (p<0.001) | +7.76% (p<0.001) |
| 50–199 | +6.72% (p<0.001) | — |
| 200+ | +3.07% (p<0.01) | −4.16% (p<0.001) |
On prior-reward alignment, the gender priors show the clearest positive correlation with observed CTR, the base score is only weakly aligned, and the Content Prior has no clear monotonic relationship. Heterogeneity is large: the Gender Prior's CTR gains concentrate in male segments, strongest for males ≤17 (+13.76%); female segments show no significant CTR effect. The Content Prior significantly lowers CTR for females 18–37 (−5.6% to −9.3%) yet lifts CVR for females 28–37 by +6.42%, so one prior can pull opposite-signed effects across groups. Title attractiveness alone explains 22% to 51% of comment-CTR variance.
The value to recommendation engineers is the pattern, not the headline number: an LLM need not be a standalone ranker or generator. It can serve purely as an offline semantic-prior estimator that warm-starts an online bandit, decoupling expensive semantic inference from low-latency serving. That division of labor is a practical template for fitting LLMs into production recommendation stacks.
For practitioners, cold-start prior design is a real problem, especially for comment and UGC settings where the text itself carries signal. The caveat is that aggregate metrics may not move at all; the gains hide in sparse-feedback and demographic slices, so evaluation has to be designed deliberately before launch.
The authors are candid. The candidate pool was manually screened (not the raw full set), so effects are conservative. The evaluation is tied to a specific discovery surface where comments sit alongside strong title visual cues; another interface may not transfer. There is no direct ablation of segment-level versus pooled posteriors, so the causal contribution of segmentation itself is not isolated. Prior alignment is clear for CTR but weak for CVR.
A few more concerns. κ=40 was chosen to make effects observable, not for production optimality; conclusions may shift under different traffic. The Content Prior's aggregate CTR is significantly negative yet is partly redeemed by a directionally positive, non-significant CVR, which is a weak "positive." Most importantly, exposure concentration: both LLM priors expose fewer unique comments and concentrate impressions on a smaller set than the uniform prior, a real cost for lightweight browsing surfaces that need freshness and rotation.