Masked Diffusion Models Mitigate the Reversal Curse via Parameter Coupling

Theoretical analysis shows that masked diffusion language models (MDMs) reduce the reversal curse because shared Transformer parameters store token-pair evidence. This parameter-level coupling allows information to transfer across positional conditionals, unlike standard autoregressive models.