ATLAS: Benchmarking and Adapting LLMs for Global Trade via Harmonized Tariff Code Classification (2509.18400v1)

Abstract: Accurate classification of products under the Harmonized Tariff Schedule (HTS) is a critical bottleneck in global trade, yet it has received little attention from the machine learning community. Misclassification can halt shipments entirely, with major postal operators suspending deliveries to the U.S. due to incomplete customs documentation. We introduce the first benchmark for HTS code classification, derived from the U.S. Customs Rulings Online Search System (CROSS). Evaluating leading LLMs, we find that our fine-tuned Atlas model (LLaMA-3.3-70B) achieves 40 percent fully correct 10-digit classifications and 57.5 percent correct 6-digit classifications, improvements of 15 points over GPT-5-Thinking and 27.5 points over Gemini-2.5-Pro-Thinking. Beyond accuracy, Atlas is roughly five times cheaper than GPT-5-Thinking and eight times cheaper than Gemini-2.5-Pro-Thinking, and can be self-hosted to guarantee data privacy in high-stakes trade and compliance workflows. While Atlas sets a strong baseline, the benchmark remains highly challenging, with only 40 percent 10-digit accuracy. By releasing both dataset and model, we aim to position HTS classification as a new community benchmark task and invite future work in retrieval, reasoning, and alignment.

• The paper introduces ATLAS, a domain-adapted LLM fine-tuned on 18,731 customs rulings to achieve 40% fully correct HTS code classifications.
• It employs supervised learning on LLaMA-3.3-70B with distributed training, ensuring robust performance on hierarchical tariff classification tasks.
• Results show significant cost efficiency at $0.70/1,000 samples, establishing a reproducible and scalable benchmark for global trade compliance.

Benchmarking and Adapting LLMs for Harmonized Tariff Code Classification: The ATLAS Model

Introduction

The paper presents ATLAS, a domain-adapted LLM for the classification of products under the Harmonized Tariff Schedule (HTS), a critical task in global trade compliance. HTS codes are hierarchical, with the first six digits harmonized internationally and the last four digits specific to the United States. Accurate assignment of these codes is essential for customs clearance, and misclassification can disrupt international logistics. Despite the importance of this task, it has been largely overlooked by the machine learning community. The authors introduce the first open-source benchmark for HTS classification, derived from the U.S. Customs Rulings Online Search System (CROSS), and demonstrate that fine-tuning LLaMA-3.3-70B yields substantial improvements in both accuracy and cost efficiency over leading proprietary LLMs.

Dataset Construction and Characteristics

The benchmark dataset is constructed by scraping and processing 18,731 legally binding rulings from CROSS, covering 2,992 unique HTS codes. Each ruling is transformed into a structured prompt-response format suitable for supervised learning, with explicit product descriptions, reasoning paths, and ground-truth HTS codes. The dataset is split into 18,254 training, 200 validation, and 200 test samples, ensuring strict separation for evaluation. The data presents unique challenges: lengthy, legalistic text, subtle product distinctions, and hierarchical correctness (6-digit vs. 10-digit). The benchmark is released on Hugging Face, enabling reproducibility and further research.

Model Training and Optimization

The authors focus on LLaMA-3.3-70B due to its dense architecture and practical deployment advantages. The model is fine-tuned using supervised learning, minimizing the token-level negative log-likelihood of the ground-truth outputs. Training is performed for five epochs (1,400 steps) with AdamW optimization, bf16 precision, and gradient accumulation to simulate a batch size of 64. Distributed training across 16 A100-80GB GPUs ensures scalability. The training loss curve exhibits rapid early improvement and stable convergence, with no signs of overfitting or instability.

Figure 1: Training loss curve over 1,400 optimization steps. Rapid early improvement is followed by stable convergence.

The authors note that dense models like LLaMA-3.3-70B are well-suited to domain-specific fine-tuning, and reproducible training is feasible under modest compute budgets. They suggest future ablations on model scale, retrieval augmentation, contrastive objectives, and direct preference optimization to further improve performance.

Evaluation: Accuracy and Cost Efficiency

The evaluation is conducted on the held-out test set of 200 samples, with three metrics: fully correct 10-digit classification, partially correct 6-digit classification, and average digit-level accuracy. ATLAS achieves 40% fully correct 10-digit classifications and 57.5% correct 6-digit classifications, outperforming GPT-5-Thinking (25% and 55.5%) and Gemini-2.5-Pro-Thinking (13.5% and 31%). The average digit-level accuracy for ATLAS is 6.3 out of 10, compared to 5.61 for GPT-5-Thinking.

ATLAS also demonstrates significant cost advantages. Inference cost per 1,000 samples is approximately $0.70 for self-hosted ATLAS, compared to$3.30 for GPT-5-Thinking and $5.50 for Gemini-2.5-Pro-Thinking. This cost efficiency, combined with the ability to self-host for data privacy, makes ATLAS suitable for deployment in sensitive trade and compliance workflows.

Implications and Future Directions

The results establish HTS classification as a challenging benchmark for LLMs, with best 10-digit accuracy at only 40%. This highlights the need for advances in retrieval-augmented reasoning, alignment, and preference optimization. The open-source release of both the dataset and the ATLAS model provides a foundation for further research in domain-specialized LLMs. The authors propose expanding the dataset, distilling ATLAS into smaller variants for edge deployment, and exploring enhanced reasoning techniques.

The practical implications are significant: improved HTS classification can reduce customs delays, enhance compliance, and increase the resilience of global supply chains. Theoretically, the benchmark exposes the limitations of current LLMs in hierarchical, high-stakes classification tasks and motivates research into more robust, interpretable, and cost-effective models.

Conclusion

ATLAS sets a new baseline for HTS code classification, achieving state-of-the-art accuracy and cost efficiency through domain-specific fine-tuning of LLaMA-3.3-70B. The open benchmark and model release invite the community to address the remaining challenges in reasoning and alignment, with direct impact on global trade and compliance. Future work should focus on dataset expansion, model distillation, and integration of retrieval and preference-based learning to further advance the field.