Goods that leave a landlocked country by land are declared again at every border they cross. A shipment moving from western China through Kazakhstan into Uzbekistan may be described in Chinese, then Russian, then Kazakh or Uzbek, and each customs office must assign it a tariff code. Trade economists find that this repeated crossing, more than raw distance, is what keeps inland economies far from markets. We propose to study an AI version of the same compounding. A language model that classifies goods correctly 95% of the time in each language does not deliver 95% clean transits. What the corridor needs is the same correct code in every language a consignment passes through. If errors in different languages are unrelated to each other, the chance of a clean transit falls with every additional border.
The core claim is that cross-lingual reasoning methods should be judged in this setting by joint correctness across the languages of a route, not by accuracy in each language separately. We also claim that methods which reason over a shared, language-independent representation should beat "translate into English, then classify" on this joint measure, even where they tie on per-language accuracy. We test this on a new parallel benchmark of customs product descriptions for the China–Central Asia corridor. We pair it with a classifier that first extracts a language-neutral record of the product's attributes and then applies the tariff rules to that record. That intermediate record is also the kind of structured, portable artifact the corridor's digital-standards literature says is missing.
Why this matters, and what prior work leaves open
Two findings motivate the proposal. Distance Puzzle: An Explanation from Global Value Chain Perspective uses gravity models, which predict trade between two countries from their size and the distance between them. It shows that the apparent rise in the effect of distance on gross trade, from an elasticity of 1.379 in 2000 to 1.430 in 2014, reverses once value-added trade is used and the number of border crossings is controlled for. Distance is costly mainly because intermediate goods cross many borders. That paper treats each crossing as a cost but does not model what happens inside one. International trade structure and analysis of complementarity between China and Poland, the Czech republic and Hungary finds that small increases in import turnover time cause large contractions in trade for landlocked Central European partners. Research on Geo-economic Cooperation between China and the Five Central Asian Countries reports that cross-border e-commerce in the region grew 95% year-on-year in 2022. It also names incompatible electronic customs standards and data formats as a core obstacle. Together these papers say the cost sits in repeated, unaligned border processing, which is the quantity our metric targets.
The closest machine-learning work is Explainable Product Classification for Customs. It uses a Korean RoBERTa model to assign Harmonized System (HS) six-digit codes, the international tariff nomenclature, across 925 hard subheadings in Chapters 84, 85 and 90. It reaches 93.98% top-3 accuracy and retrieves supporting text from the HS manual and from past dispute cases. That system works in one language, at one border, and offers no check of its decision against the rules. Customs Import Declaration Datasets is tabular only and contains no product text. Our proposal reuses the problem framing and the chapter scope of Lee et al. It departs from their work in being multilingual, in evaluating along a route, and in checking each decision against explicit rules.
On the reasoning side, there are two accounts of how multilingual models reason. Language models are multilingual chain-of-thought reasoners finds that reasoning in English beats reasoning in the native language on the MGSM math benchmark. XCOT builds on English-pivot reasoning and adds cross-lingual distillation. In contrast, The Emergence of Abstract Thought in Large Language Models Beyond Any Language finds that mature models compute through "language-shared neurons," a common internal core. In those models, tuning only the language-specific neurons gives +4.0 on MGSM for Llama-3.1-8B, against +0.3 for tuning the shared neurons. LangBridge connects a frozen multilingual encoder to an English reasoner and raises MGSM accuracy on underrepresented languages from 6.0% to 45.2%.
All of this work measures accuracy in each language on arithmetic or code. The Shi et al. paper itself flags linguistically heavy, domain-specific tasks as untested. None of these papers asks whether the errors are correlated across languages, and that is exactly what determines joint correctness. Don't Trust: Verify supplies our checking idea: translate the model's reasoning into a formal language and accept only answers that pass a mechanical check. It raised GSM8K accuracy from 67.2% (majority vote) to 79.4%. That work was English-only and limited to math. It also notes that formal provers lack libraries for regulatory text, which constrains us to a lightweight rule engine rather than a theorem prover.
Core Claim
Let a route pass through $k$ borders with languages $\ell_1,\dots,\ell_k$, and let $c_j$ be the code a model assigns to the description in $\ell_j$. We define corridor clearance rate as the probability that every $c_j$ equals the true code $c^*$. Disagreement rate is the probability that some pair $c_i \neq c_j$, which in practice triggers a dispute or reinspection.
If errors were independent across languages with per-language accuracy $a$, corridor clearance would be $a^k$. Correlated correct behavior keeps it closer to $a$. Translation pivots inject a separate translation error for each source language, so we predict their clearance falls faster with $k$ than that of methods sharing a language-neutral representation. Per-language accuracy, the number every prior paper reports, cannot distinguish the two. The falsifiable prediction is therefore that the gap between methods grows with $k$, even at matched single-language accuracy.
Research Program
The first deliverable is a parallel benchmark. We start from English and Korean product descriptions labeled with HS codes in Chapters 84, 85 and 90. Our preferred source is the dispute-case material of Lee et al., which includes roughly 200,000 international cases from 50 countries, subject to access. If that is unavailable, we use publicly published classification rulings. Professional translators render a few thousand of these descriptions into Chinese, Russian, Kazakh and Uzbek, the languages of the China–Central Asia route. Annotators write each one in declaration style rather than as a literal translation, so the versions differ in phrasing the way real filings do. A held-out subset is doubly translated to estimate how much of any disagreement comes from the translations themselves. Routes of length one to four are formed from language sequences that real consignments follow.
The second deliverable is an attribute-then-rule classifier. A multilingual model reads the description in any language and fills a fixed record: principal function, material, power source, whether it is a part or a whole machine, and whether it is a multifunction device. A small executable encoding of the HS General Interpretive Rules and the relevant chapter and section notes maps that record to a code. If the record violates a note, it is rejected and sent back to the model to re-extract. Because the rules are fixed, cross-lingual disagreement can only enter through attribute extraction. We train that extraction step for cross-lingual agreement with XCOT-style distillation: we minimize divergence between the record predicted from the English description and the records predicted from its parallel versions. For Kazakh and Uzbek we compare tuning all parameters with tuning only the language-specific neurons identified by the attribution method of Chen et al. That comparison tests whether their MGSM finding carries over to regulatory text.
The third deliverable is an analysis that connects the AI metric back to the economics. Using the crossing-count framing of Duan et al., we convert each method's disagreement rate into expected reinspections per consignment as a function of route length. We report this as a sensitivity calculation, not as a causal trade estimate.
Baselines
The naive reference point is native-language prompting: a strong open multilingual model is given the HS chapter notes in context and classifies each description directly in its own language. This shows how much consistency a general model provides with no special effort.
The strongest simple competitor is translate-then-classify. Each description is machine-translated into English and classified by the same model, fine-tuned on English HS data with the same data and the same tuning budget as our method. Shi et al. found that English-pivot reasoning is generally stronger than native reasoning, which makes this the baseline most likely to win on per-language accuracy.
The strongest existing system is the monolingual approach of Lee et al., rebuilt on a multilingual encoder and fine-tuned on all five languages. Its reported figure is 93.98% top-3 accuracy on 925 subheadings, in Korean only. For multilingual or corridor-level customs classification, no reported result was found on Lacuna. We also run a LangBridge-style frozen-encoder bridge to an English reasoner, and the XCOT recipe applied to direct classification without attribute records. On MGSM the strongest cited bridge figure is LangBridge's rise from 40.5% to 53.3% average accuracy for MetaMath-13B, and we rerun it on MGSM as a sanity check.
All systems share the same base model where architecture allows, the same training descriptions, and equal hyperparameter-search budgets. Beating a baseline means a higher corridor clearance rate at $k=3$ and $k=4$, with confidence intervals from bootstrapping over products. It also requires per-language top-1 accuracy within two points of that baseline. Winning consistency by collapsing accuracy does not count.
Two ablations isolate the key ingredients. The first removes the rule engine and lets the model output codes directly from its attribute record. The second removes the consistency distillation.
Why not just…?
Why not just translate everything to English and use the best English classifier? This is the most serious threat. If machine translation into English is good enough, it may produce consistent codes as a side effect. XCOT reports beating translation-prompt baselines on MGSM, but that is arithmetic, not tariff text, so it does not settle the question. Our route-length curve does: if translate-then-classify matches our clearance rate at $k=4$, the attribute-record idea adds nothing, and we would report that.
Why not just ensemble a large model over several translations and take the majority code? Voting raises accuracy within one border, but each border sees only its own filing, so voting does not make different offices agree. We include an ensemble baseline at matched inference cost to measure this directly.
Why not just mandate a single structured electronic declaration and skip the language problem? This policy fix is legitimate and, where it exists, better than any model. The Central Asia study reports that such standards are not interoperable across the corridor, and free-text descriptions persist in practice. Our attribute record is effectively a proposed portable format of this kind. If a hand-designed structured form filled in by traders matched our extraction step on agreement, the right conclusion would be to promote the form rather than the model. We test this with a small human-filled baseline.
Proposed Experiments and Decisive Checks
The main experiment plots corridor clearance and disagreement rates against route length, from one to four borders, for every system. Results are split into the high-resource languages (Chinese, Russian) and the lower-resource ones (Kazakh, Uzbek). The decisive result is a widening gap in favor of the attribute-then-rule classifier at matched single-language accuracy. A flat gap, or a gap favoring the translation pivot, would falsify the core claim.
A second check measures error correlation directly, as the probability that language $j$ is wrong given that language $i$ is wrong, for each system. This tests whether the mechanism works the way we claim, through shared errors, rather than through accuracy gains.
A third check repeats the language-specific versus full-tuning comparison on Kazakh and Uzbek, to see whether the shared-neuron account of Chen et al. holds on regulatory text.
A fourth check measures how often the rule engine rejects a record and how often re-extraction then corrects the code, to see whether the checking idea of Don't Trust: Verify carries over from math to tariff notes.
Finally, every model is evaluated on MGSM to confirm that the domain training has not degraded general multilingual reasoning.
References
- Distance Puzzle: An Explanation from Global Value Chain Perspective
- International trade structure and analysis of complementarity between China and Poland, the Czech republic and Hungary
- Research on Geo-economic Cooperation between China and the Five Central Asian Countries within the Framework of the Belt and Road Initiative
- Explainable Product Classification for Customs
- Customs Import Declaration Datasets
- Language models are multilingual chain-of-thought reasoners
- XCOT: Cross-lingual Instruction Tuning for Cross-lingual Chain-of-Thought Reasoning
- The Emergence of Abstract Thought in Large Language Models Beyond Any Language
- LangBridge: Multilingual Reasoning Without Multilingual Supervision
- Don't Trust: Verify - Grounding LLM Quantitative Reasoning with Autoformalization