LingDT-Decision adds information extraction to decision-model workflows

Originally reported on November 11, 2026.

Ant Digital Technologies’ LingDT-Decision adds an extract question type to the System One decision-model protocol. While a choice-based model can route a customer-service ticket to a billing team, the new feature is designed to identify information within the ticket itself—for example, the overcharged amount “199 元” in a sample complaint—and return the matching text and its position.

LingDT-Decision adds information extraction to decision-model workflows (1)

The article describes an extraction head that labels each token in the supplied text as the beginning of a span, a continuation, or outside the span. LingDT-Decision also supports multiple-choice questions. According to the article, the model uses mid-training, reinforcement learning and other techniques to adapt to enterprise tasks.

The article reports that, across 10 information-extraction benchmarks, LingDT-Decision outperformed generative models of the same or larger sizes while offering speed and inference-reliability advantages. It also reports results across 23 open benchmarks: the 27B version ranked first on a combined score weighted 90% toward single-choice and 10% toward multiple-choice tasks. These are the article’s reported comparisons, not independently verified results.

LingDT-Decision adds information extraction to decision-model workflows (2)

The model is available on Ant Digital Technologies’ DTMaaS platform, where the article says its Slotting feature can be tried through a public API example.

Original article (InfoQ China)