We describe Yandex and Yandex-8B, our unconstrained and constrained submissions to the WMT26 General Machine Translation task. Our previous WMT25 submission, a 7B translation specialist, shows strong Fluency but comparatively weaker Accuracy. To combine its translation behavior with the capacity of a much larger model without replaying the full 7B training pipeline at 235B scale, we add two forms of supervision to the instruction-tuning data of an in-house 235B pretrained model: direct translation targets generated by the 7B specialist and in-house translation-related tasks. The resulting Yandex model improves both Accuracy and Fluency and reduces the LLM-based error score relative to the same 235B model tuned only on general instructions. Relative to the 7B specialist, it improves Accuracy and reduces the error score, while their Fluency point estimates remain similar. Yandex then provides hard targets for Yandex-8B, which is trained for English-to-Russian, English-to-Belarusian, English-to-Kazakh, and English-to-Armenian using direct and instruction-conditioned translation data. Across four directions, Yandex-8B obtains the highest paragraph-level ChrF++ and the highest Accuracy and Fluency under our LLM-based evaluation among five public baselines.

LLM-based evaluators have rapidly become a standard tool for machine translation evaluation, yet the design choices that shape their behavior remain poorly understood. We present a systematic study of how instruction design, few-shot demonstrations, backbone model choice, and multi-step workflows affect span-level and ranking-based MT evaluation. We collect new expert RATE annotations on WMT24 and WMT25 English–Russian translations, with transfer validation on MQM- and ESA-annotated benchmarks across five additional language pairs. Our experiments reveal four robust findings: instruction structure shapes evaluator behavior, high-quality demonstrations matter more than quantity, frontier backbones exhibit distinct evaluator profiles (recall-, precision-, or ranking-oriented), and fusing outputs across diverse backbones outperforms single-evaluator multi-call setups. Although evaluator design exposes a large combinatorial space, several discovered design principles transfer consistently across annotation protocols and language pairs. We synthesize these findings into ORBIT, a family of LLM-based MT evaluators. Our strongest configurations achieve expert-level performance in error span detection and in Accuracy and Fluency ranking at roughly 20× lower cost than expert human annotation.

This paper describes Yandex submission to the WMT25 General Machine Translation task. We participate in English-to-Russian translation direction and propose a purely LLM-based translation model. Our training procedure comprises a training pipeline of several stages built upon YandexGPT, an in-house general-purpose LLM. In particular, firstly, we employ continual pretraining (post-pretrain) for MT task for initial adaptation to multilinguality and translation. Subsequently, we use SFT on parallel document-level corpus in the form of P-Tuning. Following SFT, we propose a novel alignment scheme of two stages, the first one being a curriculum learning with difficulty schedule and a second one - training the model for tag preservation and error correction with human post-edits as training samples. Our model achieves results comparable to human reference translations on multiple domains.