Alexandre Bonlarron

Research Focus

PhD candidate (Inria grant, Oct 2021–present) at Université Côte d'Azur, supervised by Prof. Jean‑Charles Régin. My research focuses on generating diverse, high-quality sentences for standardized vision-screening tests such as RADNER, which currently suffer from a limited pool of suitable sentences. To address this issue, we developed two Constraint Programming (CP) models that generate sentences under strict constraints. Furthermore, my current work explores how Constraint Programming can be applied to ensure that text generated by Large Language Models (LLM) adheres to predefined constraints, aiming to improve their reliability and controllability.

Key Contributions

• Introduced a CP-based decoding strategy for LLMs (backtracking + propagation).
• Propose an LLM-enhanced version of the Markov Constraint and the MDDMarkovProcess.
• Framed constrained sentence generation as a discrete optimization over MDDs.
• Generated exhaustive EN/FR sentences under strict constraints; integrated with LLMs to curate hundreds of clinically usable items.