What it is
A diffusion LM that adaptively gives harder tokens more inner-loop compute while committing easy ones early. It closes the quality gap to AR models without losing parallel decoding speed. Instead of uniform denoising depth for all masked tokens, it uses token-adaptive latent recurrence: easy tokens commit early as context, hard tokens keep refining persistent latent states. It jointly learns token prediction and per-token compute schedules via a latent-variable NELBO.
Key results
- 80.3 average over 11 benchmarks at 8B, above Qwen3-8B's 78.5
- 65.5 average at 1.7B, above Qwen3-1.7B's 63.8
- ~2.7x Qwen3-8B throughput on GSM8K at comparable or higher accuracy
Shows that matching per-token compute to prediction difficulty can make diffusion LMs both stronger than comparable AR baselines and still faster to decode.