Learned Relay Representations for Forward-Thinking Discrete Diffusion Models

In three linesLearned Relay Representations (Relay) enables Masked Diffusion Models to propagate latent information across denoising steps via a differentiable per-token channel trained with truncated BPTT. Applied to Fast-dLLM v2, it outperforms supervised finetuning on coding tasks and reduces inference latency by 32%.

Summary generated by Claude — human-verified