VGB for Masked Diffusion Model: Efficient Test-time Scaling for Reward Satisfaction and Sample Editing

Kijung JeonGeorgia Tech [email protected], Thuy-Duong VuongUCSD [email protected], Molei TaoGeorgia Tech [email protected] · 2026

Read VGB for Masked Diffusion Model: Efficient Test-time Scaling for Reward Satisfaction and Sample Editing by Kijung JeonGeorgia Tech [email protected], Thuy-Duong VuongUCSD [email protected], Molei TaoGeorgia Tech [email protected] (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2606-28301.txt.

Description

Kijung Jeon, Thuy-Duong Vuong, and Molei Tao improve masked diffusion models, generators that reveal a sequence token by token from a fully masked start. At test time, popular scaling tricks sample many candidates and keep the best, yet errors accumulate until the cost explodes. Their MDM-VGB sampler instead walks a graph of masked states and can change its mind.

Inspired by a classical backtracking chain, the sampler may unmask or remask tokens at any position, favoring moves toward higher-value partial results. That enables both reward-chasing generation and repair of weak samples. The authors prove robustness to noisy process verifiers and quadratic complexity, against the exponential cost they attribute to best-of-N heuristics. Experiments on constraint puzzles and molecular generation benchmarks support the theory with strong performance.

Catalog record

CREATORKijung JeonGeorgia Tech [email protected], Thuy-Duong VuongUCSD [email protected], Molei TaoGeorgia Tech [email protected]
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE207 KB
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2606.28301

Browse the full library · About the Tower of Babel