Reasoning models brought the biggest change in LLMs the last few years. They dramatically improve model intelligence, esp. in science, math and coding, tool use skills, reduce hallucinations and improve safety. They also pose challenges in training and their increased latency needs to be mitigated in products. I will give an overview of all these issues and focus on one more - how do reasoning models impact multilingual capabilities? I will show their potential for low-resource languages and specialized settings. I will also highlight how they require to adopt a very different approach from the standard LLMs in data gathering and fine-tuning.

Mark will give a sneak preview of recent work in this area, with some astonishing demos. The study entails (no doubt foolhardy) efforts to abstractly and operationally define intelligence, language, and understanding.

One-size-fits-all Spanish doesn’t exist. Juan and Paola reveal how Meta and RWS used AI to capture regional authenticity across Latin America.

Philipp, co-inventor of phrase-based SMT, has lived through a few revolutions in the world of Machine Translation. In his talk he will share his perspectives on the latest innovations: embracing speech and video, multilinguality in LLMs and address the question whether we are better off with LLMs or standalone models.

E-learning courses aren’t documents—they’re clicks, branches, media, variables, and timing, and that’s why existing solutions fall apart. In this talk Simon presents Articulate's journey to a purpose-built workflow for streamlined e-learning localization.

TRSB is one of the early movers implementing Quality Estimation (QE) for its Canadian customers. Ron will share the experiences from exploration to full implementation and the benefits

Embracing cultural nuances and under-resourced languages promotes diversity, autonomy, and equality in multilingual translation and makes AI more intelligent, accurate, and reliable. Ilan will describe Lexicala’s corresponding QE enhancements for language and translation models.

AI only works if the business wants to pay for it and if the training data are available. That leaves the majority of languages still in the ‘dark ages’. Jace Norton will share his perspectives on the long road for indigenous languages to travel into the future and what that means for the language workers.