Amazon AGI's ALoDLM Beats Diffusion and AR Baselines, Hits 612 tok/s at 8B
arankomatsuzaki · x · 2026-10-06
Amazon AGI researchers published ALoDLM (Adaptively Looped Diffusion Language Models), with open code and 1.7B/8B checkpoints. They attribute diffusion LMs' quality gap vs. autoregressive peers to a compute-difficulty mismatch: unknown tokens vary in predictability, yet existing DLMs apply uniform depth per denoising step. ALoDLM replaces this with token-adaptive latent recurrence—easy tokens are committed as discrete context while hard ones get extra recurrent passes. It outperforms all evaluated DLMs and matching AR baselines on average benchmark score, and a GSM8K demo shows 612.4 tok/s vs. 229.3 tok/s for vLLM-served Qwen3-8B.
Related event: Amazon's ALoDLM Speeds Up Diffusion LMs to 612 tok/s(2 posts)→
More from Research
- Ben Goertzel unveils 'Distinction-Calculus', new math for forking, merging, self-modifying agent minds — bengoertzel · 2026-10-06
- Hugging Face open-sources Repo2RLEnv: turn merged PRs into verifiable RL envs — SergioPaniego · 2026-10-06
- Agent implements Context Language Models paper itself: Hermes turns context into a managed whiteboard — Apart_Tumbleweed_303 · 2026-10-06
- Contravariance theory: for hard enough tasks, brain-DNN convergent evolution is inevitable — aran_nayebi · 2026-10-06
- CoNLL 2023 paper: instruction-tuned GPT models beat children on Theory of Mind tests — dioscuri · 2026-10-06
- How Shazam worked pre-AI: spectrogram peaks hashed into a hashtable lookup — deedydas · 2026-10-06