Ant International reports fraud model cut card losses 38% against XGBoost
Ant International says its 7-billion-parameter payment model, which reads transaction histories, transaction details and links between accounts together, cut fraud losses by about 38% in a four-month live test. The report is a preprint from the company itself, and nothing has been released for others to check.

AI illustrationWhat the study shows
Ant International, which runs the Antom merchant payment platform, describes a fraud-detection model it calls Antom 3-in-1 Transformer. It has about 7 billion parameters, the adjustable numbers a model learns. Only about 1.7 billion are used for any one task. The paper is a preprint and has not been peer reviewed.
The model combines three kinds of input. The first is the history of recent payments for the user, the card and the device. The second is a table of facts about the current payment, such as amount and merchant. The third is a graph, a map of which users are connected to which. Each input has its own component, and a final stage blends them for each prediction task.
The model is first pretrained, meaning it learns general patterns from large amounts of unlabelled payment data. It is then fine-tuned on labelled fraud cases for each business. A label here is a confirmed record of whether a payment turned out to be fraud.
The authors report a live test from June to October 2026 on 30% of platform traffic. Payments were randomly split between the new model and two existing systems. One was XGB-Online, a gradient-boosted tree model, a common type of fraud model built from many small decision trees. The other was a neural model in the style of SAINT, a published method for table data. All three worked under the same rules and could act on the riskiest 0.5% of payments.
Against XGB-Online, the authors report that the new model cut the chargeback rate by 37.74% and the overall loss rate by 37.88%. A chargeback is when a payment is reversed after a dispute, often because of fraud. Against the SAINT-style model, the loss rate fell by about a quarter.
In the first stage of the test, the models scored payments without acting on them. Among the 0.5% of payments each model ranked riskiest, the new model caught about 30% of the risky transactions. XGB-Online caught about 7% and the SAINT-style model about 17%.
Offline, the authors also report that pretraining cuts the number of labels needed. On two merchant groups, the pretrained model trained with a quarter of the labels beat the same design trained from scratch with all of them. On IBM's public but synthetic credit-card dataset, it ranked fraud better than five published transaction models, though not on every measure.
Our reading
This is a deployment report, not a recipe. It gives live business outcomes from a randomised test, which is rarer and more useful than another offline benchmark score. It also gives latency: the model answers in under 47 milliseconds on average on a single NVIDIA A10 graphics card, well inside the company's time budget for a card payment.
A concrete way to read the main result: suppose a platform can block or review 1 in 200 payments. In the authors' test, the older tree model's top picks would have caught about 7 in every 100 risky payments. The new model's picks would have caught about 30. That is the kind of gap that shows up in a loss line.
On our Readiness Scale we rate the work at level II, benchmarked. There are results on public datasets, but no released code or weights for level III. The production use is claimed by the developer itself, without independent confirmation, so we do not yet treat it as proven in production at level IV.
Read the rest of this story
Membership is free. You get every article in full and the Issue every Sunday.
We send a sign-in link, no password needed. Already a member? Sign inReadiness assessment
| Level | Status | Evidence |
|---|---|---|
| I Lab only | Met | The paper describes the model, training approach and offline experiments on internal data. |
| II Benchmarked | Met | Results on public datasets: IBM synthetic credit-card data against five published baselines, four public graph benchmarks, and TransXion, IBM CCT and MBD for fusion. Caveat: the headline comparison uses synthetic data, and the baselines are the authors' own configurations. |
| III Open source | Not yet | No code, weights or data are released. The report says architecture, training configuration and dataset construction are outside its scope. |
| IV In production | Not yet | The paper claims deployment at Ant International since June 2026 and reports a four-month A/B test, with latency figures. This is a named organisation, but the evidence is a self-reported, non-peer-reviewed account by the developer with no independent confirmation. We rate it as claimed rather than proven. |
What would move it up
Release of code, weights or at least a runnable reference version, plus full training and dataset details, would reach level 3. For level 4, independent confirmation or audit of the online results would help, ideally with confidence intervals, a longer chargeback window and a second organisation using the model.
The study
Weiyu Han, Haoyu Jiang, Jie Sheng et al.. Antom 3-in-1 Transformer: A Payment Foundation Model. arXiv, 19 November 2026. Read the paper
Notes
- Han, Jiang, Sheng et al. 'Antom 3-in-1 Transformer: A Payment Foundation Model.' Preprint, arXiv, 19 November 2026. Not peer reviewed. Source
- No code, weights or data released. The report says architecture, training configuration and dataset construction are outside its scope.
- Online test, 8 June to 14 October 2026, 30% of platform traffic. Portfolio results against XGB-Online: chargeback rate −37.74%, overall loss rate −37.88%. Against the SAINT-style model: −28.26% and −24.07%. Stage I recall in top 0.5% of scores: 29.79% (new model), 7.39% (XGB-Online), 16.52% (SAINT-style). Stage II chargeback rate against XGB-Online: −28.28% pooled, −26.09% date-weighted (Tables 1 and 2).
- IBM synthetic credit-card data, temporal test split: AP 0.3829 against 0.3412 for PRAGMA; recall at top 0.5% of 83.22% against 72.12%. PRAGMA had higher F1 (0.4091 against 0.3838) and recall at 0.1% (0.3773 against 0.3642) (Table 3).
- Model size: about 6.92 billion parameters, about 1.74 billion active per task route. Latency on one NVIDIA A10: mean 46.66 ms, P95 55.86 ms, P99 80.16 ms, within a 120 ms service budget and 250 ms end-to-end budget.
- Label efficiency: with 25% of labels, pretrained model beat full-label scratch on Merchant A (AP 0.5246% against 0.4205%; R@0.5% 8.33% against 7.81%) and Merchant B (AP 0.6928% against 0.6370%; R@0.5% 10.16% against 8.11%). AlipayHK, 732 positive training examples: 10% labels gave AP 0.8214% and R@0.5% 27.75%, against 0.1148% and 10.41% for full-label scratch.
- Ablation (Table 9): full model AP 3.89%; without sequence input 1.06%; without graph 2.94%; without tabular 2.49%.
- PRAGMA, the Revolut foundation model used as a baseline: Ostroukhov et al., 2026. Source
- Companion preprint on the sequence pretraining method: Han et al., 'TRIDENT-Seq', 2026. Source
Edited and signed by
Stefan PerdalEditor and publisher86 stories editedStefan read the study, set the readiness level and approved this story. An AI system found the paper and wrote the first draft.
How this was made
An AI model read the study and wrote this article in the structure Science Underway uses. An editor reviewed it and approved it for publication. Check the original study before you rely on the details.
How we workNo corrections

