Congrats on the ICML 2026 Outstanding Paper Award! Glad to see diffusion LLMs attracting more and more attention.
The Flexibility Trap shows that discrete diffusion can hurt reasoning under arbitrary token order: too much flexibility at the token level may skip the high-entropy forks where real reasoning decisions happen.
This is exactly we think where latent diffusion becomes exciting.
Instead of denoising arbitrary surface tokens, latent diffusion avoids the flexibility trap by utilizing continuous diffusion in latent space. This is the core idea behind LaDiR: reasoning at the semantic level, with global refinement in latent space.
In LaDiR-RL (latent diffusion RL), diffusion sampling keeps multiple latent reasoning paths alive, helping mitigate the diversity collapse that often appears when RL directly optimizes text outputs.
In our ongoing LaDiR-Multi, the same principle extends to multimodal reasoning: vision, text, code, audio, and action can enter a shared latent space, be reasoned over during denoising, and finally be discretized into text by an autoregressive model. We see very strong early results on vision-text benchmarks.
