基于能量的统一视角解读扩散模型与大语言模型

摘要

一篇论文提出用能量模型统一理解扩散模型、LLM、强化学习与推理时计算：能量函数给更理想的结果打更低分，不同学习算法可视为如何定义能量、从数据学习能量并搜索低能量解。例如自回归 LLM 可用 token 负对数概率描述，扩散模型学习将样本移向高概率区域。该解释还延伸到强化学习、规划与测试时推理，不同能量函数可组合以解决新任务而无需为每种组合单独训练模型。

摘要由模型根据原文撰写，以原文为准。

"An Energy-Based Perspective on Machine Learning"

Diffusion models, LLMs, RL, and even inference-time reasoning might all be different ways of solving the same underlying problem.

So this paper provides a unified perspective based on energy-based models, where an energy function assigns lower scores to more ideal outcomes. Different learning algorithms can then be understood through how they define this energy, learn it from data, and search for low-energy solutions.

For example, autoregressive LLMs can be described through the negative log-probabilities of tokens, while diffusion models learn how to move samples toward higher-probability regions.

The paper also applies this interpretation to RL, planning, and test-time reasoning, showing how different energy functions can be combined to solve new tasks without training a separate model for every combination.

https://www.alphaxiv.org/abs/2611.14383

基于能量的统一视角解读扩散模型与大语言模型

摘要

正文 · 原文

引用原文