HTS Code Classification Remains an Open Problem

Establish effective, high-accuracy methods for Harmonized Tariff Schedule (HTS) code classification from product descriptions, specifically assigning correct globally harmonized 6-digit codes and U.S.-specific 10-digit codes, as the task currently remains unsolved according to the Atlas benchmark results and analysis.

Background

The paper introduces Atlas, a fine-tuned LLaMA-3.3-70B model, and the first benchmark dataset for HTS code classification built from U.S. Customs Rulings (CROSS). HTS codes are hierarchical: the first six digits are globally harmonized and the full ten-digit code is required for U.S. customs compliance.

Despite Atlas achieving 40% fully correct 10-digit classifications and 57.5% at the 6-digit level, the authors explicitly state that HTS code classification remains an open problem, highlighting the need for methodological advances such as retrieval augmentation, contrastive objectives, and preference-based training. This underscores that current approaches do not yet solve the task to a satisfactory level, motivating further research.

References

These directions highlight that while Atlas establishes a strong dense-model baseline, HTS classification remains an open problem with substantial room for methodological innovation.

— ATLAS: Benchmarking and Adapting LLMs for Global Trade via Harmonized Tariff Code Classification  (2509.18400 - Yuvraj et al., 22 Sep 2025) in Subsection "Ablations and Future Work", Section "Model Training"