Seven Mental · 心智七篇
← Knowledge Atlas · Source

Energy-Based Diffusion Language Models for Text Generation

EDLM:能量扩散语言模型,非自回归文本生成
SOURCE · EDLM · NVIDIA / Stanford · ICLR 2025 · challenging the AR paradigm

Energy-Based Models × Diffusion Language Models

A residual EBM at every diffusion step — corrects the approximation error · fewer steps without degradation

MainstreamAutoregressive
t₀t₁t₂
Per-token serial · sequential dependency
ChallengerDiffusion + EBM
Parallel generation · EBM correction every step
Core innovations
Residual EBMFull-sequence evaluation at each diffusion step corrects the underlying approximation error — fixes degradation when reducing steps
Parameter sourceInitialize from a pretrained AR model, or fine-tune a bidirectional Transformer via noise-contrastive estimation
Efficient samplingParallel importance sampling · 1.3× speedup over existing diffusion models
StanceChallenges the “all LLMs are autoregressive” assumption — perplexity hasn’t caught up to AR yet, but is closing the gap. A sign of architecture diversification in LLMs.
→ diffusion-lm · autoregressive-models · energy-based-modelsresearch.nvidia.com

Energy-Based Diffusion Language Models for Text Generation

摘要

提出 Energy-Based Diffusion Language Model (EDLM),将能量模型引入离散扩散过程的每一步,在全序列级别进行评估。解决了离散扩散模型在减少采样步数时性能退化的问题。

核心思想

  • 自回归之外的范式:与逐 token 生成的自回归模型不同,扩散模型支持并行生成
  • 能量模型修正:在每个扩散步骤引入残差形式的 EBM,修正底层近似误差
  • 参数来源:可从预训练自回归模型获取,或通过噪声对比估计微调双向 Transformer
  • 高效采样:并行重要性采样算法,比现有扩散模型加速 1.3 倍

定位

这是对当前”所有 LLM 都是自回归”假设的挑战。虽然尚未达到自回归模型的困惑度,但正在接近。代表了 LLM 架构多元化的趋势。

与 Agent 工程的距离

目前与 agent 工程的直接关联有限。但如果非自回归生成范式成熟,可能根本改变 agent 系统的推理模式——例如从顺序 token 生成转向并行生成,影响 context management 和 prompt 设计的基本假设。

References

  • sources/nvidia-energy-based-diffusion-lm.md