Marianne Arriola
@mariannearr
Exact likelihood for masked diffusion is here! Gilad shows the AR-diffusion quality gap is smaller than expected
Until now, likelihood-based evaluation and policy gradient estimation for masked diffusion were difficult because the ELBO is a loose bound. Learn more below!
Until now, likelihood-based evaluation and policy gradient estimation for masked diffusion were difficult because the ELBO is a loose bound. Learn more below!
Gilad@giladturok · Mar 121/ 🚨 New paper! DUEL: Exact Likelihood for Masked Diffusion via Deterministic Unmasking
We give masked diffusion models (MDMs) proper likelihood — and therefore proper perplexity — for the first time. Turns out MDMs are closer to autoregressive models than previously thought.
We give masked diffusion models (MDMs) proper likelihood — and therefore proper perplexity — for the first time. Turns out MDMs are closer to autoregressive models than previously thought.
Open quoted post →
0 32