We describe the AMI team’s submission to the WMT 2026 General Translation shared task, participating in the English-Icelandic translation direction. Building on our 2025 submission, which relied on a 3B-parameter model and substantial rule-based post-processing, this year we aim to reduce that dependency and rely more on the translation model itself. We continually pre-train a Qwen3-4B checkpoint on a mixture of Icelandic, English, code, math, and structured data, then fine-tune it with LoRA on a retrieval-augmented, domain-conditioned instruction-tuning dataset covering five domains. At both training and inference time, source sentences are enriched with dictionary entries and back-translated example translations similar to the sentence being translated, retrieved by a dedicated RAG server, and served through a self-healing inference client that retries and validates generations. We report COMET and chrF++ scores comparing this system against last year’s submission, a larger Llama-3.1-8B model adapted with the same recipe, and an instruction-tuned Qwen3-4B baseline, alongside ablations that isolate the contribution of each part of the retrieval-augmented prompt.

We present Midheind’s system contribution to two tasks at WMT25 – Tenth Conference on Machine Translation: The General Machine Translation Task and the WMT25 Terminology Shared Task. Erlendur is a multilingual LLM-based translation system that employs a multi-stage pipeline approach, with enhancements especially for translations from English to Icelandic. We address translation quality and grammatical accuracy challenges in current LLMs through a hybrid prompt-based approach that can benefit lower-resource language pairs. In a preparatory step, the LLM analyzes the source text and extracts key terms for lookup in an English-Icelandic dictionary. The findings of the analysis and the retrieved dictionary results are then incorporated into the translation prompt. When provided with a custom glossary, the system identifies relevant terms from the glossary and incorporates them into the translation, to ensure consistency in terminology. For longer inputs, the system maintains translation consistency by providing contextual information from preceding text chunks. Lastly, Icelandic target texts are passed through our custom-developed seq2seq language correction model (Ingólfsdóttir et al., 2023), where grammatical errors are corrected. Using this hybrid method, Erlendur delivers high-quality translations, without fine-tuning. Erlendur ranked 3rd-4th overall in the General Machine Translation Task for English-Icelandic translations, achieving the highest rank amongst all systems submitted by WMT25 participants (Kocmi et al., 2025a). Notably, in the WMT25 Terminology Shared Task, Erlendur placed 3rd in Track 1 and took first place in the more demanding Track 2 (Semenov et al., 2025).