Melvin Johnson

Melvin Johnson is a vice president at Google DeepMind, where he leads Gemini post-training across reinforcement learning, reasoning, alignment, evaluation, and shared modeling recipes. His earlier work helped establish a practical way for one translation model to serve multiple languages—and translate between language pairs it had never explicitly been trained on. His subsequent research examined where that transfer succeeds, where it fails, and how shared models can learn from both speech and text.

One model across languages

Melvin joined Google in 2015 after completing a master’s degree in computer science at Stanford, where he worked with Christopher Manning in the Stanford NLP Group. At Google Research, he worked on the architecture and deployment of Google’s Neural Machine Translation system before helping develop its multilingual approach.

As first author of the collaborative paper Google’s Multilingual Neural Machine Translation System: Enabling Zero-Shot Translation, Melvin and his colleagues described a small change with substantial consequences: prepend a token specifying the desired output language. The encoder, decoder, attention mechanism, and wordpiece vocabulary could then be shared across languages without changing the underlying architecture or increasing the parameter count.

Sharing a model also allowed knowledge learned from one translation direction to help another. Beyond improving many language pairs represented in training, the system could translate between pairs it had never explicitly seen together. The paper demonstrated this zero-shot multilingual translation and investigated whether the model had learned representations of meaning shared across languages. The authors presented evidence suggesting an interlingua, rather than treating that interpretation as settled.

The work connected research with deployment. Melvin’s multilingual research developed into systems supporting more than 100 languages in Google Translate. A 2019 collaborative study described a single model handling 103 languages: transfer improved translation for low-resource languages while keeping high-resource performance comparable to competitive bilingual systems. The study also documented unresolved challenges in building a universal translation model.

Testing what transfers

Melvin’s later research broadened from translation into multilingual understanding, evaluation, and joint speech-text learning. These projects examined different limits of the shared-model approach:

• Measuring uneven language transfer. Melvin co-created XTREME, introduced in 2020 to evaluate cross-lingual generalization across 40 languages and nine tasks. Models reaching human performance on many English tasks still showed substantial gaps when transferring to other languages, particularly in syntax and sentence retrieval. Results also varied widely between languages. The benchmark made those differences visible instead of allowing English performance to stand in for multilingual capability.
• Finding when paired translations still help. In the 2021 nmT5 study, Melvin and his collaborators investigated whether parallel data—examples paired with their translations—remained useful when pretraining multilingual language models. Mixing translation objectives with language modeling improved downstream multilingual and cross-lingual performance. The gains diminished as model capacity increased, but parallel data continued to help larger models when labeled task data was scarce. The result qualified the value of explicit translation supervision: its benefit depended on model size and the data available for the eventual task.
• Learning across speech and text. Melvin coauthored Mu²SLAM, which combined unlabeled speech, unlabeled text, and supervised recognition and translation data across more than 100 languages. It used a masked-denoising objective for both speech and text, with quantized speech representations serving as prediction targets. Supervised tasks helped align the model’s representations across languages and modalities. This extended shared-model learning beyond written language while giving speech and text tasks a common training framework.

Evaluation also expanded beyond zero-shot transfer. Melvin coauthored XTREME-UP, a 2023 benchmark covering 88 under-represented languages and nine user-centered technologies, including speech recognition, optical character recognition, translation, and information access. Its premise was that small amounts of annotated data can be feasible even where language resources are scarce. It therefore tested that setting directly, rather than assuming zero-shot performance was the only relevant measure.

Leading Gemini post-training

Since 2023, Melvin has led Gemini post-training at Google DeepMind. His team works on reinforcement learning, reasoning, alignment, evaluation, and unified modeling recipes across the Gemini family.

This role follows years spent building shared multilingual models and testing how their capabilities carry across languages, tasks, and modalities. The scope has broadened to general-purpose models, with Melvin directing the post-training work that shapes how Gemini’s pretrained capabilities are expressed in its responses and actions.