Masked diffusion can do math in just 8 calls?! π€―π€―π€―
ReMaRL + frozen LLaDA-8B-Instruct on GSM8K
8 calls: 54%
16 calls: 65%
32 calls: 73%
Our policy learns to unmask and remask with GRPO. We reward only correct answers. A left-to-right bias emerges.
π Recipe & insights β
πππ
1/ The original GFlowNets paper tried PPO for sampling, but it failed. We figured out why and fixed it. And now PPO beats all objectives on standard problems including molecular graph generation in large spaces.
Continuous diffusion for text is taking off.
But most methods still diffuse in spaces that were never learned for denoising and this limits their performance.
In LDLM, we train the encoder, diffusion model, and decoder jointly, so the denoising objective shapes the representation directly.
β Not: learn a space β run diffusion on it.
β Instead: learn a diffusable space.
On OpenWebText & LM1B, LDLM achieves a better qualityβspeed trade-off than MDLM, Duo, CANDI, and FLM.
Up to 13Γ faster sampling.
π§΅ Recipe and insights below β
Latent diffusion for text.
Better quality? Faster sampling?
Seen one? Didn't think so.
Meet LDLM β beats MDLM, Duo, CANDI, FLM on OpenWebText & LM1B.
2β13Γ faster.
The trick is:
Don't run diffusion in a latent space.
LEARN a latent space FOR diffusion.
Paper: https://t.co/LYgkxTWJuE