简体中文 · English

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

Core mechanism reproduction. Defining operators run locally. Private production data, large weights and dedicated serving are not reproduced; paper results are separate.

Paper information¶

Field	Details
Paper	arXiv 2608.03382
Institution	NAVER WEBTOON
First public release	2026-08-04 (arXiv v1)
Author code	No official author code in the paper when checked on 2026-08-09.
Local adapter or method key	`llm-ts-prior`
Local reproduction code	`src/auto_research/reproductions/llm_ts_prior/`

Original paper summary¶

Background and main changes¶

Topic: LLM semantic priors for cold-start bandits. New comments have no interaction feedback but their text contains gender/content preference signals. Convert LLM semantic judgments into Beta pseudocounts and maintain Thompson posteriors by gender and age segment.

flowchart LR
    N0["Comment text"] --> N1
    N1["LLM semantic signal"] --> N2
    N2["Segmented Beta priors"] --> N3
    N3["Thompson sampling"] --> N4
    N4["Update with online feedback"]

Key figure from the paper¶

Figure 1: Overall design and components. From the paper; copyright belongs to the authors. Click for the source.

Core equations¶

\(\theta_a\sim\operatorname{Beta}(\alpha_a,\beta_a),\quad \alpha_a=1+\kappa p_a,\ \beta_a=1+\kappa(1-p_a)\)

Offline and online results¶

Four-week A/B/C with about 595,000 users per arm: Gender Prior overall CTR +1.48% (p=0.144, not significant), +9.51% in the 10–49-impression cold-start slice. Content Prior overall CTR -5.68%.

Local reproduction¶

MovieLens genres proxy text semantics. Run segmented Beta priors and 40 online Thompson updates, retaining negative and nonsignificant results.

auto-research reproduce --paper llm-ts-prior --dataset-dir data --seed 42

See stable metrics. No checkpoint is committed.

Comparison: Remove the paper-specific mechanism while matching split and budget; the experimental arm uses llm-ts-prior. Relative change is in public-seed42.json; cross-paper percentages are not comparable.

Scope¶

Local results verify execution and direction, not reproduction at paper scale.
Private features, traffic and production serving are unavailable; paper online values are references only.
Evolve-compatible architectures are registered as candidates. Serving-only methods stay runnable standalone adapters, not trainable genomes.