Research
I work on retrieval with large language models: how documents are represented so a model can find them, and how to tell when the answer built on top of them is wrong.
Directions
Generative retrieval. Instead of comparing a query embedding against an index, the model generates the identifier of the document it wants. The index lives in the parameters. This removes the separate retrieval structure, and moves the hard problem into how identifiers are constructed and decoded.
Identifier spaces. A document identifier is a discrete token sequence, and the way it is built (product quantisation, residual quantisation, hybrids of the two) decides what the model can learn to retrieve. I am interested in properties of an identifier space that predict downstream retrieval quality before training the full model.
Constrained decoding. Free generation produces identifiers that do not exist. Enforcing structural validity during decoding, through logit constraints and attention masking, keeps generation inside the index, and raises the question of what syntactic validity actually implies about semantic correctness.
Evaluation and hallucination. Retrieval pipelines fail in ways aggregate metrics hide. Building datasets and protocols that quantify hallucination and retrieval failure, rather than reporting a single score, is a prerequisite for deploying any of the above.
Students
- Alexia Allal
- CIFRE PhD, Artefact and Université d’Angers, since November 2025. Co-supervised with Sylvain Lamprier. Indexation Générative de Documents par les Modèles de Langage : Mécanismes, Robustesse et Passage à l’Échelle. How documents are encoded into discrete token sequences, which construction choices produce the best retrieval, and which training-free properties of an identifier space predict downstream quality.
- Vivien Nicolas
- PhD, IRISA / INSA Rennes, since January 2026. Co-supervised with Caio Corro and Pascale Sébillot. Constrained generation for retrieval: enforcing structural validity during decoding through logit constraints and attention masking, and what syntactic validity implies about semantic validity.
Working together
I am always interested in hearing from prospective PhD students, interns and collaborators working on retrieval, document understanding or generative models. The best starting point is an email describing what you want to work on.