PlaidQ: 0.7B continuous diffusion LM distilled to one step for code generation
AlexanderTong7 · x · 2026-09-08
A new arXiv paper introduces PlaidQ, a 0.7B continuous diffusion language model for code generation that repurposes a pretrained autoregressive model as a bidirectional denoiser over continuous token embeddings.
Key results:
- At matched scale, PlaidQ is competitive with discrete diffusion LMs on code generation
- Distillation shifts the quality-compute frontier: a 16-step student hits 31.78 and 40.49 pass@10 on HumanEval and MBPP+, beating its own teacher sampled at 512 steps
- Paired-trajectory distillation pushes to a single denoising step, achieving 7.07 pass@1 on HumanEval with functionally correct programs
The authors argue continuous diffusion offers a viable interface for few-step and one-step language generation. Paper and code are public.
More from Research
- MIT Miller Lab Amplifies Paper Arguing Bayesian Descriptions Are Not Brain Mechanisms — examachine · 2026-09-08
- NVIDIA's VoLo lands at CoRL 2026: a VLM agent that orchestrates robots through long-horizon tasks — erwincoumans · 2026-09-08
- BMVA Symposium on World Models lands in London Nov 18, with DeepMind and Microsoft speakers — CSProfKGD · 2026-09-08
- Research: task-conditioned attractors explain generalization in iterative reasoning models — burkov · 2026-09-08
- Coding agents beat hand-built data agents by 37 points with 4x fewer turns, paper finds — CShorten30 · 2026-09-08
- Gemini Pro runs research task for nearly 5 hours with barely any progress — teortaxesTex · 2026-09-08