AURORA-LM introduces a continuous-latent diffusion language model built around a high-capacity, decodable text representation. A query-based encoder-decoder produces prefix-aligned latent sequences, while a block-causal Diffusion Transformer uses flow matching to generate blocks left to right and denoise positions within each block in parallel. The method restricts only the noisy-input pathway while preserving the full clean-latent prediction target. Noise-level calibration and self-trajectory consistency address training-inference mismatch. The paper reports the strongest results among evaluated continuous and diffusion language models on OpenWebText free generation and XSum summarization, with additional gains at 1B parameters.
No heat snapshots are available in the last 24 hours.