What's this about?
Text is the last major modality still generated as discrete tokens while images, video, and audio moved to continuous latent spaces. AURORA-LM autoencodes text into a continuous latent designed for both generation and faithful decoding, then models it with diffusion (40↑).
What exactly is "discrete tokens vs. continuous latents"?
Today's LLMs pick the next word from a fixed vocabulary — discrete, one token at a time. Images use continuous latents — smooth numeric spaces you can denoise all at once with diffusion. AURORA tries to give text the same continuous treatment without losing exact word-level fidelity, the thing that broke earlier attempts.
Why does it matter?
If continuous-latent diffusion works for text, it's a genuinely different generation paradigm from the autoregressive LLMs that dominate — potentially parallel decoding and tighter multimodal unification. Long-shot research, but the kind that resets assumptions if it lands.