👋 Hi! My name is Kelvin, I have been working on ways to automatically process questions (generating, decomposing and answering ❓❓❓) for the last 5 years, from text, from knowledge graph, simple to complex.

I did my PhD (on question generation, naturally) at the Université de Lorraine in the lovely city of Nancy, France under the supervision of Claire Gardent and Thiago Castro Ferreira. My PhD was carried out under Project QUANTUM which was funded by the French National Research Agency (Agence National de Recherche, ANR), and I was part of the Synalp team (now within the MosAIk group) at the LORIA (Laboratoire lorrain de Recherche en Informatique et ses Applications) laboratory.

Recently, I’ve been working on question generation useful for discourse representations and on post-training LLMs with reinforcement learning. Some of my recent work can be found here:

     ⚒️ decomposing complex questions Here I proposed the use of panels of smaller-sized LLMs to obtain decomposition candidates as well as to select from them via voting (ranking with LLM-as-judge).

     📜 generating Questions under Discussion (QUD) Here I proposed using reinforcement learning (GRPO) to obtain improved question generation that have to meet multiple constraints requiring reasoning over a piece of discourse. QUD is an emerging linguistic framework for representing discourse structure; see here for a natural language processed-focused survey.

     🔢 captcha recognition This work is exploratory and was rapidly prototyped, and the images are in a controlled setting. Here I used GRPO with a small set of exemplars to help improve vision language model (VLM) recognition of ambiguous characters.