跳到正文
研踪PaperTrail
← 研踪 PaperTrail

arXiv:2611.26018 · v1

Does the Workspace Do the Work? A Causal Protocol for Verifying Global-Workspace Computation in a Language Model

首次提交 2026-11-19 20:33
版本更新 2026-11-19 20:33
雷达观测 2026-11-20 16:53
arXiv ↗PDF ↗

摘要

Global Workspace Theory (GWT), a leading theory of consciousness, holds that information becomes conscious when it is broadcast to specialized processes that use it. Building a workspace into a network does not reveal what the trained network computes with it: a module that reports an answer may have combined a broadcast message with its own knowledge, or merely forwarded an answer computed elsewhere, and task accuracy and ablation cannot tell these apart. We introduce a causal protocol that identifies who computes when modules communicate. Its central test transplants a message from another input, chosen so that ignoring, relaying and recombining it predict three different answers. On networks built to relay or to recombine, the test assigns the right computation in every case. Among trained workspace models it finds all three outcomes: one leaves its broadcast nearly unused, and of two models that differ only in their training data, one relays and the other recombines. In AFGW, a language model whose modules communicate only through a learned workspace, a transplanted message yields the answer that recombination predicts in 97% of cases, and its content acts by redirecting what the receiver retrieves from its own memory. The tests carry over to other communication designs and to a known interface of a pretrained Transformer, turning GWT's central claim into a testable property of trained models.

作者备注:39 pages, 8 figures

AI 论文解读

基于摘要 · 非全文审阅

此版本尚未生成 AI 解读。解读由后台任务生成并按论文版本保存,登录后共享已有结果。

版本差异

目前只收录了一个版本。收录下一版后,可以对照标题、作者、分类和摘要。

观测到的版本

相邻记录可能跳过中间版本;首次发现 v2 及以上也会标为版本更新。

v1 · 2026-11-19 20:33 · 最新观测版本

Does the Workspace Do the Work? A Causal Protocol for Verifying Global-Workspace Computation in a Language Model

Global Workspace Theory (GWT), a leading theory of consciousness, holds that information becomes conscious when it is broadcast to specialized processes that use it. Building a workspace into a network does not reveal what the trained network computes with it: a module that reports an answer may have combined a broadcast message with its own knowledge, or merely forwarded an answer computed elsewhere, and task accuracy and ablation cannot tell these apart. We introduce a causal protocol that identifies who computes when modules communicate. Its central test transplants a message from another input, chosen so that ignoring, relaying and recombining it predict three different answers. On networks built to relay or to recombine, the test assigns the right computation in every case. Among trained workspace models it finds all three outcomes: one leaves its broadcast nearly unused, and of two models that differ only in their training data, one relays and the other recombines. In AFGW, a language model whose modules communicate only through a learned workspace, a transplanted message yields the answer that recombination predicts in 97% of cases, and its content acts by redirecting what the receiver retrieves from its own memory. The tests carry over to other communication designs and to a known interface of a pretrained Transformer, turning GWT's central claim into a testable property of trained models.

查看此版本解读 →查看此版本原文 ↗