PlaidQ：单步扩散式代码生成

PlaidQ: Writing Code in One Step — The Diffusion Model Revolution Duke University and Tsinghua University's Breakthrough in Code Generation Published: September 10, 2026 | Reading time: 10 minutes The Revolutionary Research On September 3, 2026, researchers from Duke University and Tsinghua University published a groundbreaking paper that answers a fundamental question in AI code generation: Can language models write code using diffusion models — and do it in just one step? The answer is yes. What Is PlaidQ? PlaidQ is a continuous (Gaussian) latent-diffusion language model that works differently from traditional autoregressive models: Traditional LLMs (Autoregressive) Generate tokens left to right Each token depends on all previous tokens Sequential process — slow for long sequences PlaidQ (Diffusion) Diffuses a whole sequence in a 16-dimensional continuous token-embedding latent Denoises with a bidirectional Qwen3-0.6B trunk Can generate all tokens simultaneously — or in just a few steps The Distillation Breakthrough The key innovation is distillation — reducing the number of denoising steps: Steps Description Performance 512 Original diffusion process Baseline 16 Distilled to 16 steps Student outperforms teacher on HumanEval pass@10 1 Distilled to single step Can generate executable code, but HumanEval pass@1 is only 7.07 The 16-step model demonstrates that students can surpass teachers on certain benchmarks. The 1-step model shows the feasibility of parallel generation, though it's not yet reliable for high-quality coding. Code Example: Using PlaidQ import torch from plaidq import PlaidQModel # Load the distilled model model = PlaidQModel.from_pretrained("plaidq-0.7b") # Generate code in a single step prompt = """ def fibonacci(n): """Generate Fibonacci sequence.""" if n <= 1: return n return fibonacci(n-1) + fibonacci(n-2) """ # PlaidQ generates the entire sequence at once result = model.generate(prompt, num_steps=16) print(result) # Or even in one step (experimental) result_one_step = model.generate(prompt, num_steps=1) print(result_one_step) Enter fullscreen mode Exit fullscreen mode Why This Matters For Code Generation Speed: Generate entire code sequences in parallel, not sequentially Efficiency: Fewer steps mean faster generation Quality: 16-step model outperforms teacher on certain benchmarks For AI Research Diffusion Models: Show potential beyond image generation Distillation: Demonstrate effective knowledge transfer Parallel Generation: Challenge the autoregressive paradigm For Developers Faster Iteration: Generate code faster than traditional LLMs Better Quality: 16-step model achieves high pass rates New Paradigm: Explore diffusion-based code generation Performance Comparison Model HumanEval pass@1 HumanEval pass@10 MBPP pass@1 Teacher (512 steps) 65.85 78.05 72.30 Student (16 steps) 63.41 80.73 70.15 Student (1 step) 7.07 15.30 8.20 Key Insight: The 16-step student outperforms the teacher on HumanEval pass@10, demonstrating effective distillation. The 1-step model shows feasibility but needs improvement. Technical Details Architecture Backbone: Qwen3-0.6B (bidirectional trunk) Latent Space: 16-dimensional continuous token embeddings Diffusion Process: Gaussian noise addition and removal Distillation: Knowledge transfer from 512-step to 16-step model Training Data: Code datasets (HumanEval, MBPP) Method: Distilled continuous diffusion Goal: Reduce steps while maintaining quality Code Example: Distillation Process from plaidq.distill import distill_model # Load teacher model teacher = PlaidQModel.from_pretrained("plaidq-teacher") # Distill to student model student = distill_model( teacher=teacher, num_steps=16, dataset="humaneval", epochs=10 ) # Evaluate student score = student.evaluate("humaneval", metric="pass@10") print(f"Student pass@10: {score}") # Distill further to 1 step student_1step = distill_model( teacher=student, num_steps=1, dataset="humaneval", epochs=5 ) Enter fullscreen mode Exit fullscreen mode Future Directions Short-term Improve 1-step model quality Extend to more code benchmarks Optimize distillation process Long-term Apply to other domains (text, images) Combine with autoregressive models Develop hybrid generation methods Conclusion PlaidQ represents a significant step toward parallel code generation using diffusion models. The ability to generate code in just 16 steps — or even 1 step — challenges the traditional autoregressive paradigm and opens new possibilities for AI code generation. While the 1-step model is not yet reliable for production use, the 16-step model demonstrates that students can surpass teachers through effective distillation. This research highlights the potential of diffusion-based language models and the importance of distillation in achieving high-quality, efficient code generation. This article is based on research published by Duke University and Tsinghua University on September 3, 2026. Paper: arXiv:2609.04531 | Code: github.com/pengzhangzhi/plaidq