本文探讨了一种新型的非神经元架构，旨在解决大型语言模型的信息瓶颈问题。作者提出了一种基于词嵌入的技术，通过将单词表示为向量，并在模型各层中使用特定数量的神经元生成单词，从而构建一个类似备忘录的机制。该备忘录可以被解释、转述，并与CoT（思维链）一起加载到模型中，用于后续步骤。此外，文章还提出了另一种方法，即通过多个神经元组生成单词并置于准CoT中，训练模型使它们保持一致，并允许在适当的时候清除它们。文章还强调了这种架构在可解释性方面的优势，以及其在计算效率上的潜力。

🧠**核心问题：** 现有模型在处理长序列信息时面临信息瓶颈，导致模型难以理解和利用大量信息。

💡**解决方案：** 作者提出了一种非神经元架构，利用词嵌入技术将单词转化为向量，并通过模型各层中的特定神经元生成单词，构建一个类似备忘录的机制。

📝**备忘录机制：** 备忘录的长度等于模型的层数，其中包含模型生成的单词。备忘录可以被解释、转述，并与CoT一起加载到模型中，用于后续步骤。

🔄**另一种方法：** 另一种方法是使用多个神经元组生成单词并置于准CoT中，训练模型使它们保持一致，并允许在适当的时候清除它们。

✅**优势与潜力：** 这种架构在可解释性方面具有优势，并且在计算效率上具有潜力，有助于更好地理解模型的工作原理。

Published on May 20, 2025 11:35 PM GMT

Neuralese recurrence, allowing a model to learn thousands of tokens of information from a single pass, is thought to be the key to building high-capability models. Unfortunately, this key lacks interpretability, ensuring that we cannot check the alignment of the model by looking at its thoughts.

Back in April 3 a user lacking formal computer science education proposed an idea of a non-neuralese architecture something along the line "maybe to solve this information bottleneck problem, we should solve the generate-many-tokens-in-one-pass problem." Unfortunately, he didn't propose any technical solutions. With this post I aim to propose the solution with which I came up. 

There is a technique of word embedding which represents words as vectors and even allows one to draw sense-faithful conclusions like "queen ≈ king-man+woman".  The most primitive such representations have dimension 100-300, which allows one to define a word by using only 300 neurons. Meanwhile, a layer consists of tens of thousands of neurons. Consider a setup where the first 300 neurons of each layer generate a word to be put into the memo which will have length equal to the number of layers, while the model is to train to produce the memo that will resemble a text on a subject similar to the problem. Then the memo might be interpreted, paraphrased^[1] to avoid euphemisms and attempts to communicate in an alien language^[2] and loaded into the model along with the CoT for the next step. 

Alternatively, one could use the approach where the words are generated by many groups of neurons and placed into quasi-CoTs; then the model should be given access to all the quasi-CoTs, trained to make them coherent and have the option to purge them when the time comes.

 

1. ^{^}

   GPT-4o estimates the amount of compute necessary to paraphrase a 100-token memo by GPT-2 as ~5 billion floating point operations. The neural network with many parameters needs far more work than the paraphraser. 

2. ^{^}

   Although the model may develop a foreign language, it will be immune to the paraphraser and easier to understand given access to the memos and their context.

Discuss