ALoDLM: Adaptively Looped Diffusion Language Models
paperYour notes
Diffusion language models denoise every token to the same depth, which spends the same compute on a token that was settled after one pass as on one that is still contested. ALoDLM makes the depth a property of the token: it "iteratively refines latent representations and allocates computation according to token difficulty," committing tokens that are ready into discrete context while the unresolved ones keep cycling through further recurrent passes. The per-token schedule is treated as a latent variable and optimised through a conditional NELBO, so the allocation is learned rather than heuristic.
The reported result is the one that matters for the class: at both 1.7B and 8B it beats every evaluated diffusion LM and the corresponding autoregressive baselines on average across eleven benchmarks, while keeping the parallel decoding that is diffusion's reason for existing. Diffusion LMs have generally traded quality for decode speed; the claim here is that adaptive depth recovers the quality without giving the speed back. Weights ship at both scales under CC BY-NC 4.0 — non-commercial, so this is a research release rather than a deployable one.
Thirteen authors led by Liancheng Fang, with weights published from Amazon's own HuggingFace organisation and code under amazon-science; the arXiv abstract page carries no affiliation block, so the lab attribution rests on those two. This is the clearest frontier-technique output Amazon shipped in the window, and it clears the scale floor: the 8B demonstration is the one that makes it relevant to how anyone would actually build a model.