Credit: Microsoft

Microsoft has launched Adaptive Custom Translation (AdaptCT) in Microsoft Foundry Tools, a major upgrade to how organizations build domain-specific translation systems. Instead of retraining custom models from scratch, AdaptCT uses existing LLMs such as GPT-4o and GPT-4o-mini together with tiny datasets and smart indexing to deliver instant, high-quality translations tailored to each business.

Solving the limits of traditional custom translation

Conventional custom translation requires massive datasets, lengthy training cycles, and repeated retraining for every update. This slows down teams managing fast-changing content like support tickets, product releases, or documentation.

AdaptCT eliminates these blockers. Businesses can now create domain-specific translation behavior with as little as five aligned sentence pairs, scaling up as needed—no retraining, no deployment steps, and no maintenance overhead.

Why AdaptCT matters

AdaptCT gives organizations the speed and flexibility needed for global-scale communication:

• Updates in minutes by refreshing a dataset and rebuilding an index

• Support for 5–10,000 pairs, ideal for niche domains

• No model retraining required

• Lower cost and faster iteration

• Direct integration with Foundry and Translator APIs

The result is translation that stays accurate as terminology evolves, without the operational burden of traditional systems.

Real-world impact

Early users have already seen measurable wins:

• Customer support: Launch multilingual feature updates with 50–500 sentence pairs instead of 10,000+.

• E-commerce: Maintain consistency across catalog updates with datasets that evolve daily.

• Tech documentation: Keep translations aligned with frequent API changes using lightweight term indexes.

AdaptCT turns translation into a strategic enabler rather than a bottleneck.

How it works

AdaptCT retrieves the most relevant aligned sentences from an indexed dataset and uses few-shot learning to adapt the LLM’s translation output. The result blends general language fluency with domain-specific terminology and style.

Developers can get started by:

1. Creating a project

2. Uploading aligned sentence pairs (in TMX or TSV)

3. Building an index

4. Calling translate APIs with the dataset index

AdaptCT supports translation to/from English and requires a GPT-4o deployment in Azure AI Foundry.

When to use AdaptCT vs. Custom Translator

AdaptCT is ideal for:

• Low-volume or frequently changing content

• Domains with limited data

• Support tickets, chat logs, product updates

• Scenarios requiring rapid iteration

Custom Translator remains the choice for:

• High-volume, stable content

• Legal/medical use cases with strict terminology

• Large datasets (10,000+ pairs)

Best practices

Microsoft recommends:

• Start with 50–100 high-quality sentence pairs

• Organize indexes by domain

• Version datasets as terminology evolves

• Use blind test sets before deployment

• Provide full-sentence context, not isolated terms

A new era for translation customization

Adaptive Custom Translation marks a shift toward flexible, data-efficient translation systems powered by LLMs. The approach cuts development cycles, reduces cost, and helps enterprises deliver real-time multilingual experiences with up-to-date terminology.

Developers can explore samples on GitHub, watch the demo on YouTube, and get started through the Azure AI Translator documentation and the Microsoft Foundry portal.