Learning to Re-Draft: A Variational Stackelberg Game for Discrete Diffusion
VSDD learns a corruption process so discrete diffusion models can re-draft tokens more effectively.
Variational Stackelberg Discrete Diffusion learns which token substitutions are most useful for training a discrete diffusion denoiser to re-draft earlier tokens. Training is a leader-follower game: the leader defines a Markovian corruption process from the denoiser embeddings, and the follower optimizes a variational denoising objective with that process fixed. Improvement is measured under a reference corruption process, approximated with a one-step gradient update, and optimized with a score-function estimator. On molecular, text, and playlist generation, VSDD improved molecular validity over uniform and masked diffusion, lowered text perplexity versus uniform diffusion, and improved offline playlist recommendation metrics.
- Learns a semantically aware corruption process instead of fixed masking or uniform noise.
- Cast as a Stackelberg game between a corruption leader and a denoiser follower.
- Leader rewards corruptions by later denoiser improvement, not easy reconstruction.
- Improves molecular validity, text perplexity versus uniform diffusion, and playlist metrics.
- AI models
- VSDD
From the original article
1 min read at arxiv.orgDiscrete diffusion models offer the ability to re-draft, revisiting and correcting earlier tokens throughout generation. This capability depends on the forward corruption process that defines what the denoiser learns to correct. Masked diffusion models fix tokens once they are unmasked, while uniform diffusion permits revisions but relies on uniformly random token substitutions.
By Dmitrii Moor · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.