InfluMatch uses 4B models for low-cost Thai influencer matching

The authors present InfluMatch, a three-stage cascade using small open-weight models to match influencers to multi-part Thai marketing criteria efficiently.

• Dense retrieval returns 50 candidates, which are scored by a 4B reranker keeping the top 10.
• A 4B reasoner grades the shortlist per criterion with a Thai rationale.
• The cascade reaches 94.1% P@5 on an 11-query set, matching Kimi-K2.6 accuracy.
• It emits ~35x fewer output tokens and serves queries in ~20s on one A100.
• Pairwise fine-tuning of the reranker improves accuracy, while reasoner fine-tuning degrades end-to-end ranking.

The system provides a deployable, explainable KOL search solution at a small fraction of frontier serving cost.