In today's market, matching KOLs (Key Opinion Leaders) to multi-part Thai marketing criteria typically relies on keyword-based structured profile searches, which often miss semantic fits; using frontier LLMs to evaluate candidates one by one, while accurate, is expensive and slow. To address this, we present InfluMatch, a low-cost three-stage cascade: retrieval $\rightarrow$ rerank $\rightarrow$ reason, built entirely from small open-weight models.
The process works as follows: first, dense retrieval returns 50 candidates; then, a 4B pointwise reranker scores each by the log-probability of a single 'Yes' token and retains 10; finally, a 4B reasoner grades the shortlisted candidates per criterion on a rubric with Thai rationale. This cascade is designed for cost efficiency: reasoning over a filtered top-10 halves token spend compared to reasoning over all 50, while scoring 14 points higher.
In end-to-end comparison against human relevance labels on an 11-query set, the full cascade achieves 94.1% P@5, while a retrieval-only baseline is nearly random; it matches the frontier model Kimi-K2.6 (91.8%) while emitting approximately 35 times fewer output tokens and servicing a 50-KOL query in about 20 seconds on one A100. Notably, the only fine-tuning that pays off is pairwise: a SimPO-tuned reranker matches the frontier baseline's best-pick accuracy (78.0 EM), whereas fine-tuning the reasoner on pointwise per-criterion labels improves offline scores but degrades end-to-end ranking—an inversion traced back to the design of the absolute labeling task, leaving the untuned base model as the strongest deployed reasoner. The result is a deployable, explainable KOL search system at a fraction of the frontier serving cost.
Blogger's Review: InfluMatch successfully reduces the cost of KOL search through a cascade design of small open-weight models while maintaining high matching accuracy. This innovative approach not only has significant technical implications but also offers new insights for applications in marketing. Its unique perspective on fine-tuning methods provides valuable references for future model optimization.