Parallel Language Models paper: multi-token generation at AR quality, up to 5.7x speedup
aaron_lou · x · 2026-10-30
An arXiv paper introduces Parallel Language Models (PLMs), which generate mutually consistent tokens in one forward pass by routing dependencies through latent auxiliary variables, trained with a variational bound (AR loss + parallel decoding loss). PLMs match near-AR pretraining loss at 80M–3B scale, with speedup following a power law in model size. A post-trained 1.3B model matches or exceeds an identically trained AR model on math, coding and instruction-following while speeding up inference 2.3–5.7x, sitting on a better speed-quality Pareto frontier than AR and discrete diffusion baselines. Code will be open-sourced.
More from Infra
- YC-backed VirajAero launches turbine that powers datacenters with 40% more power from same fuel — Scobleizer · 2026-10-30
- Proof of Agent rolls out 15-minute checkpoints anchored to Solana and Base, capping validation costs — adamamcbride · 2026-10-30
- Pooling $8K for a shared local LLM rig takes 3.3 years to beat a $20/mo ChatGPT sub — Any-Blacksmith-2054 · 2026-10-30
- 10 coding agents tested on local Qwen 3.8 27B: Pi wins on efficiency at 9.2k tokens per task — latent_lurker8587 · 2026-10-30
- Banks Shelve $5-7B Junk-Bond Sale for Anthropic-Leased Texas Data Center on Grid Audit — rohanpaul_ai · 2026-10-30
- Weight compression on the GPU: researcher proves his own instinct wrong — realDanFu · 2026-10-30