Back to Blogs
Nov 5, 2026Blog

Research Paper: Delivering Accurate Translations across Devices with Adaptive Models

Research Paper: Delivering Accurate Translations across Devices with Adaptive Models

We're excited to share that our research paper, "EDWARD-NMT: Elastic Dimensions With Adaptive Routing and Dual-consistency for Neural Machine Translation" has been accepted at EMNLP 2026 - one of the leading international conferences in Natural Language Processing (NLP). Our research addresses a practical challenge in AI translation: How can we deliver accurate translations across different devices without requiring large, expensive AI models?

The Challenge: Making AI Translation Accessible Everywhere

While AI translation has improved significantly, high-quality models often require substantial memory and computing power, making them difficult to run on smartphones and resource-limited devices. Building separate models for different devices is a common solution, but it adds extra training, maintenance, and infrastructure costs.

At UNEY, where we develop secure technologies like end-to-end encrypted (E2EE) chat applications, efficient on-device translation enables users to communicate across languages while keeping their conversations private.

EDWARD-NMT: Train Once, Deploy Anywhere

EDWARD-NMT introduces a new approach built around one simple idea: train one AI model that can adapt to different devices and workloads. Think of it as a translation engine with adjustable capacity. On a powerful server, it can use its full capabilities for maximum accuracy. On a smartphone, it can operate in a smaller, lighter configuration without needing to be retrained.

EDWARD-NMT achieves this through three key ideas:

  • Flexible Model Size: One trained model can be transformed into smaller versions to fit different devices, without additional training.

  • Smart Computation: Instead of using the same computing power for every word, the model dynamically decides how much processing is needed.

  • Consistent Translation Quality: A new training technique helps smaller versions preserve the translation quality of their larger counterparts.

Together, these innovations make AI translation more flexible, efficient, and practical for real-world applications.

Research Results

We evaluated EDWARD-NMT across English–Vietnamese, English–German, and 16-language multilingual benchmarks, demonstrating competitive or superior translation accuracy over established models with lower computational cost. Notably, all variants derive from a single trained model without additional fine-tuning, with the smallest version running 2.2× faster on-device using roughly 4× fewer parameters than the full-size model.

What This Means for UNEY

EDWARD-NMT reflects our commitment to building AI technologies that are both powerful and practical. For UNEY's end-to-end encrypted communication applications, this research opens up opportunities for private, on-device multilingual translation, allowing people to communicate across languages without relying on external translation servers. Beyond messaging, the same technology can support multilingual communication services across smartphones, desktops, and cloud platforms, adapting to the resources available.

Our goal is simple: Make language barriers smaller without making AI models bigger.

Read the paper: EDWARD-NMT: Elastic Dimensions With Adaptive Routing and Dual-consistency for Neural Machine Translation

Get in touch

We would love to hear from you

Reach out to discuss how UNEY can help secure your digital world.

Email us