← News

ResearchSources: arXiv · 18 Nov 2026

Visa Research fine-tunes a 1B Llama model to top tabular classifiers

The team says VRLLM, a fine-tuned Llama 3.2 1B, beat a range of tabular classifiers and their own production deep learning model on large-scale tabular data.

In plain words

Tabular data means rows of structured values, like transaction records, rather than text. Visa Research reports that a small pretrained language model, after fine-tuning on billions of labeled rows, classified such data better than established tabular models. The claim comes from the paper's abstract, not an independent evaluation.

What was announced

The arXiv abstract states that prior work often found LLMs lagging behind tree-based models and purpose-built neural architectures on tabular classification. The Visa Research team fine-tuned a pretrained Llama 3.2 1B on billions of rows of labeled tabular data, naming the result VRLLM. They report it outperformed a wide array of tabular classifiers and their competitive production deep learning model.

The paper describes the data preparation, training recipe, ablation studies and deployment validation behind the result, and also benchmarks frontier LLMs on the same task. Interpretability experiments suggest VRLLM learns feature-specific predictive associations, and component ablations point to the earliest two FFN layers as contributors to linearly decodable task information. The authors position the work as an engineering recipe and empirical case study for adapting compact LLMs to industrial tabular prediction.

What it does not show

  • The evidence is the paper's abstract only; the full paper's methods, datasets and results were not reviewed, and no independent confirmation of the performance claims is available.

For security and IT teams

Scope
The reported outperformance of tabular classifiers and the production deep learning model is Visa Research's own claim about its setting, described in the abstract as deployment validation for an industrial tabular prediction task.

Sources

Source-support review completed 19 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.

Coverage: arXiv: abstract.

Updates and corrections

No corrections so far.

First published 19 Nov 2026

New to the subject? AI from the Ground Up starts from the beginning.