08/05/2026
The debate over whether to use LLMs for translation is over. The real conversation now is how to deploy them without losing quality, control, or security. 🌟🌍
That the number one insight our AI research team came back with from EAMT 2026, the European Association for Machine Translation's annual conference.
We're beyond proud to share that six of our researchers had papers accepted at this year's conference.
📄 Marina Sánchez-Torrón, Daria Akselrod, and Jason Rauchwerk co-authored a paper titled, "To Write or to Automate Linguistic Prompts, That Is the Question." ❔🤔 They found that AI-written instructions worked almost as well as ones written by expert linguists, except at catching translation errors, where the humans still won.
📄 Olivia Norris and Alex Yanishevsky collaborated on a paper titled, "LLM-as-a-Jury for Machine Translation Publishability Assessment." Instead of asking one AI model whether a translation is ready to publish, this team asked several at once — and the group vote beat any single model. 🤝
📄 Vitalii Iakivchuk's paper, titled, "Using Model Disagreement to Identify Unstable Regions in MT Evaluation," which found that when human reviewers disagree on a translation's quality, it's actually a good thing. That disagreement is useful because it points to exactly where evaluation gets shaky. 👀
We just published a full recap of the team's time at the conference, plus what we're changing about our own approach because of it, on our blog.
Read it here: https://bit.ly/4xlG9OK