Sticky Jump Diffusions: A Unifying View of Masked, Continuous, and Hybrid Diffusion

Score Entropy Discrete Diffusion (SEDD) parameterizes discrete reverse processes with unconstrained positive score ratios. While positivity guarantees nonnegative reverse jump rates, it does not ensure Bayes realizability: ratios at a noisy state need not be jointly induced by any clean-token posterior under the forward kernel. The score-entropy loss has the correct population optimum but does not enforce this constraint away from it. In a trained pure-uniform SEDD checkpoint, roughly one quarter of complete score vectors violate the coordinate box, while more than half lie inside it yet remain materially incompatible with any valid posterior. Such violations can produce negative pre-normalization weights in finite-step sampling. Projecting raw scores onto the bridge polytope removes all observed negative weights and improves external generative PPL from $203.6$ to $175.1$ without changing the sampler. We introduce \emph{mean-to-score} (M2S), which predicts a clean-token posterior mean and converts it to the score through an exact kernel-dependent linear map. The construction applies to any known coordinate-wise continuous-time Markov chain (CTMC) satisfying a mild support condition. For uniform corruption, it maps the probability simplex onto the bridge polytope; for absorbing-mask corruption, the resulting objective recovers MD4 exactly. In a controlled 28.4M-parameter CIFAR-10 comparison, M2S lowers test BPD from $3.173$ to $3.129$ and FID-50k from $\CifarSEDDFID$ to $\CifarMtwoSFID$. A 170M-parameter M2S model trained on about 262B OpenWebText token slots outperforms the evaluated pure-uniform SEDD, GIDD, and Neural CTMC checkpoints at every tested sampling budget, reaching generative PPL $143.3$ at 128 steps versus $183.6$ for the strongest pure-uniform baseline.

Representation alignment with pretrained vision models has recently shown strong potential for accelerating diffusion transformer training. By aligning intermediate diffusion features with clean-image representations from self-supervised vision encoders, existing methods improve convergence and generation quality. However, such alignment also introduces a non-trivial constraint: diffusion models operate on noisy inputs whose usable information varies across timesteps, while the reference features are extracted from clean images. In this paper, we revisit this mismatch from a token-level perspective. We find that, under full-token representation alignment, tokens with large alignment-gradient norms exhibit a stable spatial preference, suggesting that the alignment objective does not affect all tokens uniformly and may encourage the model to rely on the complete set of clean-image tokens. To address this issue, we propose MaskAlign, a token-subset representation alignment method that applies alignment to randomly sampled token subsets during training. By exposing the model to different token subsets across iterations, MaskAlign reduces the dependence of representation alignment on the complete token set and encourages alignment behavior that is more stable under token-subset perturbations. To mitigate the information loss caused by directly dropping tokens, we further introduce a lightweight pre-mask token mixing block that shares information across tokens before masking.

Masked diffusion models (MDMs) generate discrete sequences by iterative denoising under an absorbing masking process. In standard masked diffusion, if a token remains masked after a reverse update, the model discards its clean-state prediction for that position. Thus, still-masked positions must be repeatedly inferred from the mask token alone. This design choice limits cross-step refinement. To address this limitation, this paper proposes a simple, yet effective, post-training adaptation for MDMs that conditions each denoising step on the model's own previous clean-state predictions. The resulting method, called Self-Conditioned Masked Diffusion Models (SCMDM), requires minimal architectural change, does not introduce a recurrent latent-state pathway, does not rely on an auxiliary reference model, and adds no extra denoiser evaluations during sampling. This is an important departure from partial self-conditioning approaches which requires expensive model training from scratch. In particular, the paper shows that partial self-conditioning, including the commonly used 50% dropout strategy for training self-conditioned models from scratch, is suboptimal in the post-training regime. Instead, once the model's self-generated clean-state estimates become informative, the specialization to refinement is preferable to mixing conditional and unconditional objectives. SCMDM is evaluated across multiple domains, demonstrating consistent improvement over vanilla MDM baselines, achieving nearly a 50% reduction in generative perplexity on OWT-trained models (42.89 to 23.72), alongside strong improvements in discretized image synthesis quality, small molecular generation, and enhanced fidelity in genomic distribution modeling.