Vincent-Pierre Berges

Learning Facts at Scale with Active Reading

HybridCoT: Interleaving Latent and Text Chain-of-Thought for Efficient Reasoning

Learning to Reason for Factuality

Continual Learning via Sparse Memory Finetuning

Memory Layers at Scale

LOCATE 3D: Real-World Object Localization via Self-Supervised Learning in 3D

From Thousands to Billions: 3D Visual Language Grounding via Render-Supervised Distillation from 2D VLMs