Carving Thought from Noise: A Deep Dive into Diffusion Language Models

*Full English adaptation of a Chinese forum post surveying the rise of diffusion-based language models.*

The Giant Bound by Chains: The Iron Law of Autoregression

Since the Transformer architecture appeared in 2017, the autoregressive (AR) paradigm has dominated NLP. GPT, LLaMA, Claude, and Gemini all generate text like a typewriter—left to right, one token at a time, each token conditioned on all previous ones.

This paradigm has a deep-rooted limitation: it can only look forward, never back. Every decision is hostage to prior content, and the model lacks the ability to revise earlier choices during generation. Ask an AR model to write a sentence backwards, and it struggles; fill-in-the-blank tasks that require using both left and right context also expose its weakness.

Meanwhile, a wave from the image-generation world has been quietly rolling toward the shores of language.

The Magic in the Darkroom: How Diffusion Models Work

Diffusion models power Midjourney, DALL-E, and Stable Diffusion. Inspired by physical diffusion, the forward process gradually adds noise to data until it becomes pure random signal; the reverse process trains a neural network to denoise step by step, reconstructing or creating new data.

The core obstacle for language: images are continuous pixel values, while text is a sequence of discrete symbols. There is no "middle state" between the words "apple" and "banana." Bridging this gap required new ideas.

Bridging Two Worlds: Three Technical Routes

| Route | Representative models | Core idea | Strengths / limits | |---|---|---|---| | Continuous-space DLMs | Diffusion-LM, TESS 2 | Map tokens to continuous embeddings, diffuse, map back | Inherits mature image-diffusion techniques, but suffers semantic loss in translation | | Discrete-space DLMs (masked) | D3PM, LLaDA, Dream | Replace tokens with [MASK] via transition matrices, then recover | No semantic loss; now the mainstream route | | Riemannian DLMs | RDLM (KAIST) | Map discrete text into a curved continuous space via Riemannian geometry | Smooth diffusion paths, avoids abrupt semantic "jumps" |

In masked diffusion models, the [MASK] token—familiar from BERT's cloze tasks—plays the role that noise plays in image diffusion: generation becomes the gradual resolution of a fully masked sequence into meaningful text.

The Challenger to the Throne: LLaDA

In 2024, researchers at Renmin University of China's Gaoling School of AI introduced LLaDA (Large Language Diffusion with mAsking)—the first diffusion LLM capable of genuinely fluent dialogue. The name deliberately echoes "llama."

LLaDA's core idea in one sentence: replace token-by-token continuation with bidirectional context demasking. Training masks random spans of the response given a prompt; generation starts from a fully masked answer and progressively resolves it over multiple denoising rounds—like a sculptor who views the whole marble block before cutting, rather than chiseling left to right.

At 8B parameters, LLaDA matches LLaMA3-8B on many mainstream benchmarks, leads it by 5 percentage points on GSM8K, and even surpasses GPT-4o on reversed-language tasks (writing sentences backwards, reversing letter order).

A Constellation of Models (2024–2025)

| Model | Institution | Size | Highlights | |---|---|---|---| | LLaDA-8B | Renmin University | 8B | Masked diffusion pioneer; GSM8K +5% over LLaMA3-8B | | LLaDA-MoE | Renmin University | Sparse MoE | Activating only 1.4B params matches dense Qwen2.5-3B | | Dream-7B | University of Hong Kong | 7B | Beats both LLaDA-8B and LLaMA3-8B on multiple benchmarks | | Mercury | Inception Labs | undisclosed | Thousands of tokens/sec; GPT-4-class performance | | Gemini Diffusion | Google DeepMind | undisclosed | Google's first industrial diffusion LLM | | RDLM | KAIST | — | Riemannian geometry for smooth discrete diffusion |

The Parallel Awakening: Speed and Inference Acceleration

Autoregressive generation is inherently serial—one forward pass per token, like a single-lane highway. Diffusion models can predict all masked positions simultaneously per denoising step. Key accelerations:

• Step distillation: compressing dozens of denoising steps into a few or even one.
• FreeCache: porting KV-cache ideas to semi-autoregressive diffusion decoding for up to 34× speedup.
• Semi-autoregressive decoding (e.g., SSD-LM): parallel within blocks, serial across blocks—a practical compromise.

Can Diffusion "Think"?

Diffusion models historically meshed poorly with chain-of-thought (CoT) reasoning, since parallel generation clashes with ordered inference steps. Recent work is closing the gap:

• Diffusion-of-Thought (DoT): treats intermediate "thoughts" as noisy diffusion states inferred in parallel with mutual constraints, converging into a coherent reasoning chain.
• Diffu-GRPO and VRPO: transplant policy-gradient and preference-optimization (RLHF-style) training to diffusion LLMs, yielding notable gains in math reasoning and code generation.

Multimodal Fusion

• MMaDA: a unified masked diffusion model that simultaneously beats LLaMA3 on text reasoning, Show-o on multimodal understanding, and SDXL on text-to-image generation—in a single architecture.
• LLaDA-V extends LLaDA to vision-language tasks.
• ThinkDiff couples a VLM's "reasoning brain" with a diffusion model's "creative hands," enabling visual inference followed by image creation.
• DIFFA: the first diffusion-based large audio-language model, beating several AR baselines on spoken-language understanding benchmarks.

Real-World Applications

• Code generation (DiffuCoder): bidirectional context suits infilling—given a function's head and tail, it fills the middle, outperforming left-to-right-only approaches.
• Summarization (InfoDiffusion): high-information tokens are resolved early in denoising, mirroring how humans draft main ideas before details.
• Agents: parallel planning lets models draft multiple action paths at once; full-context bidirectional generation improves cross-turn coherence; Mercury-class speeds satisfy real-time interaction.

Remaining Hard Problems

1. Length flexibility: non-autoregressive parallel generation requires fixing sequence length up front—like knowing how many crossword cells exist before filling them. Semi-autoregressive schemes are the main workaround; natural "stopping" remains an open problem. 2. Quantization: denoising produces large activation outliers that degrade low-bit quantization more severely than in AR models, hindering edge deployment. 3. Infrastructure debt: a decade of software (PyTorch, vLLM, TensorRT) and hardware optimization has centered on AR models; diffusion lacks mature inference engines, so theoretical speed advantages are partly eroded in practice.

What Comes Next

• Scaling laws: whether diffusion LLMs sustain gains—and whether their parameter-performance curve is steeper than AR—remains to be verified.
• Unified multimodality: diffusion's shared denoising framework makes it a natural backbone for true unified multimodal models.
• Deeper RL integration: efficient alignment algorithms for diffusion are a frontier.
• Native hardware support: purpose-built kernels and engines could rapidly erase today's tooling gap.

The likely endgame is not one paradigm annihilating the other, but division of labor: autoregressive models keep dominating streaming conversation, while diffusion models shine in code infilling, document drafting, creative writing, and multi-agent systems where global planning matters.

Epilogue

In 1944, Schrödinger wrote that *life feeds on negative entropy*. Diffusion language models are a digital echo of that idea: from a fully masked, chaotic starting point, they draw in the "order" of the prompt and gradually sculpt meaning out of noise. The autoregressive model is a traveler moving forward in time, each step defined by the past; the diffusion model is a meditating poet who first gazes at the whole world, then—slowly, holistically—awakens a poem from chaos. As AI's future unfolds, we are really asking: should a machine think more like flowing water, or more like sculpture?

References

1. Nie, S., Zhu, F., Du, C., et al. (2025). *LLaDA: Large Language Diffusion with Masking.* arXiv:2502.09992. 2. Du, Y., Li, S., Torralba, A., et al. (2023). *Diffusion-of-Thought: Chain-of-Thought Reasoning in Diffusion Language Models.* arXiv:2402.07754. 3. Zhao, Z., Li, M., Huang, Z., et al. (2025). *MMaDA: Multimodal Large Diffusion Language Models.* arXiv preprint. 4. Ye, J., Guo, Z., Li, Y., et al. (2025). *Dream: Efficient Diffusion Language Model with Reasoning Enhancement.* arXiv preprint. 5. Austin, J., Johnson, D. D., Ho, J., et al. (2021). *Structured Denoising Diffusion Models in Discrete State-Spaces (D3PM).* NeurIPS 2021.