ALoDLM: Adaptively Looped Diffusion Language Models
ALoDLM gives harder-to-predict tokens more compute while letting easier ones commit sooner.
The paper frames diffusion language models’ quality gap as a mismatch between uniform denoising compute and uneven token difficulty. ALoDLM adds token-adaptive latent recurrence, keeping unresolved tokens in refinement while feeding committed tokens back as discrete context. The authors train 1.7B and 8B models and report higher average scores than evaluated diffusion models and corresponding autoregressive baselines across eleven benchmarks. They also claim the model keeps fast parallel decoding under optimized inference engines. Source: HF Daily Papers' note
The paper frames diffusion language models’ quality gap as a mismatch between uniform denoising compute and uneven token difficulty. ALoDLM adds token-adaptive latent recurrence, keeping unresolved tokens in refinement while feeding committed tokens back as discrete context. The authors train 1.7B and 8B models and report higher average scores than evaluated diffusion models and corresponding autoregressive baselines across eleven benchmarks. They also claim the model keeps fast parallel decoding under optimized inference engines. Source: HF Daily Papers' note
score 5