Researcher, Meta AI Research
2 papers at NeurIPS 2025
Transition matching (TM) is a discrete-time continuous-state generative modeling that advances both flow/diffusion and autoregressive models. TM variants achieve state-of-the-art text-to-image generation.
We introduce Resample-Previous-Tokens (RPT), a sampling method that allows models to revisit and replace previously generated tokens, leading to ~10% relative improvements in reasoning and coding after a short fine-tuning.
