Ant Digital's LingDT-Decision-1.0: Industrial AI Decision Model with Native Multi-Select & Sequential DAG

Model Overview

On October 22, Ant Digital Technology (蚂蚁数科) released LingDT-Decision-1.0, a decision model purpose-built for rapid industrial decision-making. The model comes in two sizes — 4B and 27B parameters — and is designed to handle structured decision tasks with higher accuracy, lower latency, and drastically reduced token consumption compared to general-purpose large language models (LLMs). Unlike LLMs that generate text token-by-token and then require parsing, LingDT-Decision-1.0 directly selects from a predefined set of options, eliminating format instability and simplifying validation pipelines.

Core Innovation: Native Multi-Select & Sequential DAG Decisions

Most existing decision models support only single-label classification or scoring. LingDT-Decision-1.0 introduces two industry-first capabilities:

Native Multi-Select : A single inference can hit multiple candidates simultaneously, fitting multi-label recognition, compound intent detection, and tool-combination selection.

Sequential Decision (DAG) : The model respects dependency ordering among tasks — e.g., in a user request "refund, cancel auto-renewal, reissue invoice, but do not close account," it can deduce that refund verification must precede invoice processing while auto-renewal cancellation can run in parallel.

These capabilities mirror real enterprise workflows where a single interaction triggers multiple dependent actions, reducing missed requirements and improving end-to-end accuracy.

Benchmark Performance (23 Open-Source Benchmarks)

Evaluation uses a composite Decision Capability Index = 0.9 × Single-Select Score + 0.1 × Multi-Select Score.

Single-Select (13 benchmarks): JEVal, OpenDecision, SemIf-144, Scienthoon-pooled, WANLI-256, TypeSafe-102, Nimble-282, JevBench-231, BANKING77-3080, Transfer-v9, Decision-v7, Transfer-v4, MASSIVE.

Multi-Select (10 benchmarks): SATA, MultiRC, Jigsaw, BlendX, BRIGHTER-en, BRIGHTER-zh, GoEmotions, ECtHR-A, ECtHR-B, UNFAIR-ToS.

Results: The 27B version ranks first overall, surpassing Microsoft-Decision-1, Jev 1.13.0, Fastino/GLiDE, Star-lux, and 12 other decision models. In multi-select tasks, both 4B and 27B versions lead all 16 compared models, demonstrating a clear advantage in composite judgment. In single-select, they sit in the top tier.

Real-World Deployment Results

Ant Digital tested the model in three production domains:

1. Data Annotation (vs. DS-V4-flash)

Token cost reduced ~97%

Processing latency improved ~10×

2. Intelligent Customer Service (vs. Qwen-3.7-flash)

Token cost reduced ~90%

Latency improved ~1× (approx. 2× faster)

Risk identification accuracy rose from 76% to 89%

3. Financial Applications (vs. Qwen-8B generative model)

Latency improved ~5×

Tool-combination exact-match rate increased ~2.6×

Hallucination recall improved from 44% to 70%

Across scenarios, the model consistently delivers lower cost, faster response, and higher judgment accuracy while a single model adapts to multiple decision types, simplifying integration and maintenance.

Architectural Shift: From Generation to Direct Judgment

Traditional LLM pipelines:

User Input → LLM generates structured text → Parser extracts decision → Validator checks format → Retry on failure

. LingDT-Decision-1.0 replaces this with:

User Input + Candidate Set → Model directly outputs selected option(s) in predefined order

. This removes the generation-parsing-validation loop, cutting token usage and latency while stabilizing output format.

Availability

LingDT-Decision-1.0 is live on the Ant Digital DTMaaS platform: https://maas.antdigital.com/models