第七章 · Section 7
Working Memory (WM) · 工作记忆
Working Memory (WM): The ability to maintain, manipulate, and update information in active attention.
工作记忆（WM）：在活跃注意力中保持、操作与更新信息的能力。
文本工作记忆
Textual Working Memory
Recall: Remember a short sequence of elements.
Sample: [Fleep, Zorp, Glim, Chair] — "State the nonsense words in alphabetical order."
Transformation Sequence: Remember and update a short list of digits.
Sample: [10, 20, 30] — "First, append the number 40. Then reverse the list."
回忆：记住一个短元素序列。
样例：[Fleep, Zorp, Glim, Chair]——"按字母顺序说出这些无意义词。"
变换序列：记住并更新一个短数字列表。
样例：[10, 20, 30]——"先加上数字 40，然后反转列表。"
听觉工作记忆
Auditory Working Memory
Recall: Remember a collection of sounds or voices.
Transformation Sequence: Remember and modify a short utterance.
Sample: "Say 'the brown fox jumps over the dog.' Now say it with a deeper voice and make it sound like a question."
Tone Sequence Comparison: "Listen to these tone sequences: [C4, E4, G4, F4, A4] and [C4, E4, F4, G4, A4]. Are they the same?"
回忆：记住一组声音或人声。
变换序列：记住并修改一个短语句。
样例："说'the brown fox jumps over the dog'。现在用更低沉的声音说，并让它听起来像问句。"
音调序列比较："听这两段音调序列：[C4, E4, G4, F4, A4] 和 [C4, E4, F4, G4, A4]。它们相同吗？"
视觉工作记忆
Visual Working Memory
Recall: Remember a collection of images.
Sample: "Which plane in (B) was also in (A), if any?"
Transformation Sequence: Transform a visual input.
Sample: "Finish the sketch."
Spatial Navigation: Represent a sense of location.
Sample: "If I'm facing the kitchen window, is the refrigerator to my left or right?"
Long Video Q&A: Understand a long video or movie.
Sample: "After watching the movie Wicked, who took credit for levitating Nessarose?"
回忆：记住一组图像。
样例："(B) 中的哪架飞机也出现在 (A) 中（如果有）？"
变换序列：变换视觉输入。
样例："完成这幅草图。"
空间导航：表征环境中的位置感。
样例："如果我正对着厨房窗户，冰箱在我的左边还是右边？"
长视频问答：理解长视频或电影。
样例："看完电影《魔法坏女巫》（Wicked）后，谁因让 Nessarose 悬浮而居功？"
跨模态工作记忆
Cross-Modal Working Memory
Cross-Modal Association: Remember cross-modal correspondences.
Sample: "Which animal corresponds to 'dog'?"
Dual N-Back: Monitor visual and audio streams and detect matches over time (Dual n-back test).
跨模态联想：记住跨模态的对应关系。
样例："哪个动物对应'dog'？"
双任务 N-back：持续监测视觉与听觉流并检测匹配（双 n-back 测验）。
评估细节与 AI 表现
Assessment Details & AI Performance
Assessment Details. See Appendix E for further details on how to assess working memory capabilities concretely.
AI System Performance. The table summarizes current AI system performance on Working Memory (WM) tasks. While the raw Textual Working Memory score appears similar between GPT-4 and GPT-5 in this battery, improvements in managing long contexts are also reflected in the Document Level Reading Comprehension score within the Reading and Writing (RW) ability.
Model | Textual (2%) | Auditory (2%) | Visual (4%) | Cross-Modal (2%) | Total
GPT-4 | 2% | 0% | 0% | 0% | 2%
GPT-5 | 2% | 0% | 1% | 1% | 4%
评估细节：如何在具体层面评估工作记忆，参见附录 E。
AI 系统表现：下表汇总当前 AI 系统在工作记忆（WM）任务上的表现。虽然本套测验中 GPT-4 与 GPT-5 的原始"文本工作记忆"分数看似相近，但管理长上下文的改进也反映在读写（RW）能力中的"文档级阅读理解"分数里。
模型 | 文本(2%) | 听觉(2%) | 视觉(4%) | 跨模态(2%) | 合计
GPT-4 | 2% | 0% | 0% | 0% | 2%
GPT-5 | 2% | 0% | 1% | 1% | 4%