alchaincyf/global-workspace-paper-zh

This repository is a complete Chinese translation of an Anthropic research paper called Verbalizable Representations Form a Global Workspace in Language Models, published in July 2026 on Anthropic's Transformer Circuits research site. The paper describes a discovery made by Anthropic's interpretability team while studying their Claude models. They found a structure inside the model, which the translator calls a workspace, that resembles what a person might call conscious awareness: information the model can report on, hold onto, and reason about, sitting above a much larger layer of computation the model cannot put into words. The paper also introduces a tool for reading and adjusting this workspace, along with two practical uses, one for checking whether a model is behaving honestly and one for training the model to reflect on hypothetical situations. The repository itself does not contain any code to run. Instead it holds the translated material in two forms, an 85 page formatted PDF with all nine main chapters and about fifty figures, and a plain markdown file with the same text. There is also a folder with high resolution copies of all 94 figures from the original paper. According to the README, the appendix was not translated, and the translator explains upfront how mathematical formulas and citations were handled. The translation was produced with AI assistance and then checked by a person credited as Hua Shu, who also writes commentary about the paper on a WeChat public account. The README is clear that copyright belongs to Anthropic, this is an unofficial translation meant for learning and discussion, and readers should treat the original English paper as authoritative. Anyone who spots a translation error is invited to open an issue. This is best suited for Chinese speaking readers who want to study Anthropic's interpretability research without working through the original English text, rather than developers looking for a tool or library to install.