ALoDLM spends more compute on the hard tokens

Amazon AGI has published ALoDLM, a diffusion language model that decides how much computation each token gets instead of applying the same depth everywhere. The work is described at alo-dlm.github.io, with models at 1.7B and 8B parameters and code released alongside.

Diffusion language models generate by predicting many tokens at once, which is what makes them fast, but they have trailed comparably sized autoregressive models on quality. ALoDLM's authors put that gap down to a computation–difficulty mismatch: inside one denoising step, some masked positions are easy to predict and others need far more work, yet a standard denoiser spends the same fixed depth on all of them. Confidence-based decoding only defers the hard ones, and a deferred token's latent computation is discarded and recomputed from its mask embedding at the next step.

A shared core that runs up to four times per step

ALoDLM keeps parallel denoising and adds an inner loop. A Prelude encodes the partially masked sequence once per denoising step; a shared Recurrent Core then runs up to K = 4 times, with a Coda readout after each pass feeding an LM head and an exit gate. After every pass, unresolved positions whose predictive entropy falls at or below a threshold are sampled and committed, and committed tokens re-enter the next pass as token embeddings. Positions that are not ready keep their latent states and refine them. The loop ends when the mean cumulative halting probability over unresolved positions reaches a threshold, when nothing is unresolved, or at the fourth pass. Latent states persist only within a step.

Training treats the per-token exit schedules as latent variables and derives a conditional negative evidence lower bound, so token prediction and computation allocation are learned together. The authors write that the model learns token-dependent refinement preferences without explicit difficulty labels.

612.4 tokens per second against 229.3

ALoDLM-8B is described as the largest looped diffusion language model to date. Across eleven reasoning, coding and knowledge benchmarks it has the best average score among the evaluated diffusion models and beats its Qwen3 autoregressive counterparts: 65.5 against 63.8 at 1.7B, and 80.3 against 78.5 at 8B, after supervised fine-tuning on a 5B-token corpus.

On GSM8K, ALoDLM-8B reaches 612.4 tokens per second against 229.3 for vLLM-served Qwen3-8B, roughly 2.7 times the throughput at comparable or higher accuracy. Recurrent depth also works as a test-time scaling axis: more passes buy more quality.