课堂-模型原理与论文证据边界Lesson - Model Mechanics and Paper Evidence Limits

能从摘要分出研究问题、证据位置、实验条件三栏。

能写出至少一条作者没有主张的内容，并说明依据来自原文哪一句。

能用自回归生成与注意力解释输出为什么逐词产生，不说成查表或理解。

能区分「窗口装得下」与「窗口里的信息被用上」，并引用位置效应的结论。

Separate an abstract into research question, evidence location and experimental conditions.

Write at least one thing the authors did not claim and point to the original sentence supporting that boundary.

Use autoregressive generation and attention to explain why output is produced token by token, without calling it lookup or understanding.

Distinguish information fitting in the window from information in the window being used, citing position-effect findings.

arXiv 1706.03762《Attention Is All You Need》摘要页：翻译任务、BLEU 与训练设置。

arXiv 2307.03172《Lost in the Middle》摘要页：多文档问答与键值检索中的位置效应。

本站《论文卡-Attention Is All You Need》作为读法示例。

三栏工作表：主张、证据位置、条件与未主张。

arXiv 1706.03762 Attention Is All You Need abstract: translation tasks, BLEU and training setup.

arXiv 2307.03172 Lost in the Middle abstract: position effects in multi-document QA and key-value retrieval.

This site’s paper card for Attention Is All You Need as a reading example.

A three-column worksheet: claim, evidence location, conditions and non-claims.

1. 3 分钟读第一篇：圈出任务名与指标，写出它主张的范围。
2. 3 分钟读第二篇：把结论写成条件句，说明信息位置变化时表现怎么变。
3. 3 分钟互查：小组交换工作表，把无条件断言改写成带条件的转述。
4. 3 分钟连回机制：用自回归与注意力解释这两篇的观察，指出哪些说法证据不足。

1. 3-minute first paper: circle task names and metrics and write the scope of the claim.
2. 3-minute second paper: write conclusions as conditional sentences and describe how performance changes with information position.
3. 3-minute peer check: groups exchange worksheets and rewrite unconditional claims as conditional paraphrases.
4. 3-minute mechanism link: use autoregression and attention to explain the observations and identify statements with insufficient evidence.

把标题或摘要里的强措辞当结论；结论受任务与设置限制。

把长窗口当成「长文档都能答对」；信息在中间时表现可能下降。

把注意力当成人类式「注意到重点」；它是权重计算。

把某个指标提升当成能力全面升级；BLEU 对应特定翻译任务。

Treating strong wording in a title or abstract as the conclusion; conclusions are bounded by task and setup.

Treating a long window as able to answer every long document correctly; performance can drop when information is in the middle.

Treating attention as human-like noticing the key point; it is a weight computation.

Treating one metric gain as a general capability upgrade; BLEU corresponds to a specific translation task.

三栏工作表填完，「未主张」一栏至少一条能回连原文。

转述句含任务、设置与观察结果，不含「理解」「总是」这类无条件词。

能说明本课没有复跑实验，也不给出自测分数。

The worksheet is complete, and at least one non-claim can be linked back to the original text.

Paraphrases include task, setup and observed result, without unconditional words such as understands or always.

Learners can state that this lesson did not rerun experiments and gives no self-test score.