Researchers have developed a deterministic agentic workflow for Harmonized System (HS) tariff classification, a complex task requiring multi-dimensional rule reasoning. Unlike self-planning agents, this fixed control flow confines language model calls to specific stages, ensuring interpretability through structured outputs and verbatim citations of relevant notes. When evaluated on the HSCodeComp dataset using a Qwen3.6-plus model, the workflow achieved significant accuracy at both four- and six-digit levels. An open-weight Qwen3.6-27B-FP8 backbone also demonstrated strong performance, and a manual audit suggested potential discrepancies in existing ground-truth labels. AI
Summary written by gemini-2.5-flash-lite from 1 source. How we write summaries →
IMPACT Introduces a structured approach to complex rule-based AI tasks, potentially improving accuracy and interpretability in specialized domains.
RANK_REASON Academic paper detailing a new methodology for AI-driven classification. [lever_c_demoted from research: ic=1 ai=1.0]