@InProceedings{steingrmsson-rarson-daason:2026:wmt,
  author    = {Steingrímsson, Steinþór  and  Þórðarson, Sveinbjörn  and  Daðason, Jón Friðrik},
  title     = {What a DRAG (it is being small): The AMI Submission to the WMT 2026 General Translation Shared Task},
  booktitle      = {Proceedings of the Eleventh Conference on Machine Translation},
  month          = {October},
  year           = {2026},
  address        = {Budapest, Hungary},
  publisher      = {Association for Computational Linguistics},
  pages     = {1421--1431},
  abstract  = {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.},
  url       = {https://aclanthology.org/2026.wmt-1.79}
}