We propose a method to improve machine translation quality by ensembling multiple large language models within a multi-agent framework. The method addresses the problem of ensembling models with different tokenizations by abstracting away from token-level representations and operating at the level of generated text. The approach integrates several established concepts in MT, including ensembling, quality estimation, and reranking into one translation system. The log-likelihood reranking applied in this paper significantly improves results, even without ensembling. Experimental results show improvements comparable to COMET-MBR.

SRPOL team submission to WMT2025 introduces innovative approach using A* (A-star) algorithm of decoding in EuroLLM which gives diverse set of translation hypotheses. Subsequent reranking by Comet-QE and NLLB chooses the best of the diversed hypotheses which gives significant improvement of translation quality. The A* algorithm can be applied to decoding in any LLMs or other translation models. The experiment shows that by using free, openly accessible MT models you can achieve translation quality of the best online translators and LLMs using just a PC under your desk.

This paper presents the system description of Samsung R&D Institute Poland participation in WMT 2022 for General MT solution for medium and low resource languages: Russian and Croatian. Our approach combines iterative noised/tagged back-translation and iterative distillation. We investigated different monolingual resources and compared their influence on final translations. We used available BERT-likemodels for text classification and for extracting domains of texts. Then we prepared an ensemble of NMT models adapted to multiple domains. Finally we attempted to predict ensemble weight vectors from the BERT-based domain classifications for individual sentences. Our final trained models reached quality comparable to best online translators using only limited constrained resources during training.