Timothée Lacroix

Mistral’s summit biography identifies Lacroix as a 2015 computer-science graduate of ENS rue d’Ulm and a holder of the Mathematics, Vision, Learning master’s from Paris-Saclay. It describes thesis work from 2016 to 2019 in collaboration with École des Ponts on tensor factorization for recommender systems.[2]

Lacroix began as an engineer at Facebook AI Research in New York in 2015, according to his employer biography. He continued at Meta with Guillaume Lample before co-founding Mistral. Mistral identifies him as CTO and describes his responsibility as leading its engineering teams. The public sources establish that broad responsibility, while the papers provide a more specific account of the research to which he contributed.[1][2]

Finding missing relationships

His 2018 paper with Nicolas Usunier and Guillaume Obozinski studies knowledge-base completion: estimating missing relationships in a collection of facts. It represents that collection as a multidimensional array and learns a compact decomposition. The paper shows that how training is constrained and how inverse relationships are represented can materially change results. Its experiments improved a comparatively simple decomposition method, illustrating that the setup of a comparison can matter as much as choosing a more elaborate model.[7]

Adding time to facts

A later paper with Obozinski and Usunier extends the problem to facts whose validity changes over time. A relationship in a social network or a person’s interaction with an item is not necessarily permanent. Their method adds a time dimension and introduces training constraints suited to this setting. The work also supplies a Wikidata-derived evaluation dataset. It offers a concrete example of why learning from a static list of facts can miss information needed to answer a dated question.[6]

Efficient language models

Lacroix co-authored the original LLaMA paper, which describes models trained on publicly available data, and later the Mistral 7B and Mixtral reports. LLaMA’s results compare model size and performance under the paper’s evaluation setup. Mixtral uses a router to choose a small subset of expert blocks for each token. These reports are the clearest starting points for his language-model work, but their author lists do not assign each architecture decision to an individual.[3][4][5]