Gustave Cortal
I'm a research scientist at Callyope (opens in a new tab), doing natural language processing for mental health. I'm building a safety benchmark for language models. I also analyze spontaneous speech for psychiatric disorders (opens in a new tab), summarize clinical documents, and optimize inference for clinical workloads.
During my PhD in natural language processing (opens in a new tab) at ENS Paris-Saclay (opens in a new tab), I developed language models to analyze subjective experience (opens in a new tab) and connect emotion analysis with cognitive science theories (opens in a new tab).
My personal projects explore interpretability (thought steering), transformer-based models for niche domains (psychotherapy, dreams, animal motion) and image generation (ganlive, imagevolve). Most are optimized for efficient inference in the browser.
Latest writing
Steering LLM thoughts with cognitive reframing
Subtracting the activation difference between reframes and negative thoughts makes the thoughts a language model writes self-reproachful, and adding it makes them dismissive. The layer sets what changes.
News
- Oct 2026I released five web demos: a tiny transformer of animal motion (animacy), activation steering for reframing (thought steering), a recommender for rare concepts (dérive), a GAN latent space explorer (ganlive) and text-guided image evolution (imagevolve).
- Aug 2026Our paper on interpretable topic modeling of spontaneous speech in depression (opens in a new tab) was published in JMIR Mental Health.
- Jan 2026I defended my PhD, Natural language processing for subjectivity analysis in personal narratives (opens in a new tab), at ENS Paris-Saclay.
- Nov 2025Our paper Formalizing Style in Personal Narratives (opens in a new tab) was presented at EMNLP.
- Oct 2025I released Piaget (opens in a new tab) and Beck (opens in a new tab), language models fine-tuned on psychological reasoning traces, then aligned on psychotherapeutic preferences with ORPO.