Clément DUMAS

Education 📚

• 2024-2025: Computer Science research Master MVA (Mathematics, Vision, Learning) at ENS Paris-Saclay. The ENS is a selective institution that trains teachers and researchers
• 2023-2024: Computer Science research Master MPRI at ENS Paris-Saclay.
• 2022-2023: Double Bachelor's degree in Computer Science at ENS Paris-Saclay
• 2020-2022: Completed "classes préparatoires", an intensive two-year programme in the sciences with 12 hours of math per week, preparing for the competitive entrance exams to the ENS

Research Experience 🔬 (✨ = led to a first-author publication)

• Since September 2025: Mentoring 5 junior researchers in the SPAR program to work on 2 model diffing projects: understanding model personas and disentangling fine-tuning objectives
• July 2025: Published a blog post explaining our motivation for working on model diffing, our paper, follow-up analysis and next steps
• Since May 2025 ✨: MATS extension in Neel Nanda's stream. I was awarded a 1-year scholarship to continue my model diffing research. We published our paper Overcoming Sparsity Artifacts in Crosscoders to Interpret Chat-Tuning at NeurIPS 2025. Our follow-up work "Narrow Finetuning Leaves Clearly Readable Traces in the Activation Differences" was spotlight at the NeurIPS 2025 mechanistic interpretability workshop
• Winter 2025: MATS 7.0 scholar in Neel Nanda's stream working on model diffing. An early version of our paper was published in the ICLR SLLMs workshop
• October 2024: Completed the Neel Nanda's MATS stream training phase. It ended with a 2 weeks research sprint where we replicated and extended the Crosscoder paper. See our demo Colab
• Summer 2024 ✨: 5-month research internship at EPFL with Robert West and Chris Wendler. Our work, Separating Tongue from Thought: Activation Patching Reveals Language-Agnostic Concept Representations in Transformers , was spotlight at ICML 2024 mechanistic interpretability workshop and accepted at ACL 2025 main conference
• July 2024: Attended the Human-aligned AI Summer School
• January 2024: Explored the emergence of XOR features in LLMs and the RAX hypothesis developed by Sam Marks.
• October 2023 - May 2024: Supervised Program for Alignment Research (SPAR) under the supervision of Walter Laurito. We tried to apply Contrast-Consistent Search to Reinforcement Learning models .
• Summer 2023: Two months research internship with Jobst Heitzig on Aspiration-Based Q-Learning
• 2022-2023: Participated in “Séminaire Turing”, an AI alignment reading group at ENS Paris-Saclay
• December 2022: Participated in the AI testing hackathon organized by Apart Research. Our submission about Trojans in transformers was ranked #4
• November 2022: Participated in the Interpretability hackathon organized by Apart Research
• November 2022: Participated in the ML4G, a one-week French AI alignment camp organized by Effisciences
• October 2022: Participated in the AI alignment Hackathon organized by EffiSciences
• 2021-2022: Implemented local playout optimization in a Monte-Carlo tree search for the travelling salesman problem

Programming Projects 🛠️

• Developed mechanistic interpretability tooling: nnterp (✨ poster at the NeurIPS 2025 mechanistic interpretability workshop), a wrapper around NNsight focused on LLMs and tiny-dashboard, a tool to visualize activations of SAEs
• Early adopter of the mechanistic interpretability library NNsight, actively engaging with the community to answer questions and improve it
• Developed a CodinGame multiplayer game
• Ranked in top percentiles in CodinGame multiplayer bot programming contests: top 0.5%, 3% and 7% in 2021-2022
• Proficient in OCaml, Python, Java, PyTorch, and NNsight, with working knowledge of Rust, CUDA, C++ and others

Referees