SPG uses upper and lower bounds on log-likelihood to provide a better policy gradient for RL in diffusion LLMs, outperforming ELBO-based methods on math and puzzle tasks.
Title resolution pending
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.CL 1years
2025 1verdicts
UNVERDICTED 1representative citing papers
citing papers explorer
-
SPG: Sandwiched Policy Gradient for Masked Diffusion Language Models
SPG uses upper and lower bounds on log-likelihood to provide a better policy gradient for RL in diffusion LLMs, outperforming ELBO-based methods on math and puzzle tasks.