Common Mistakes

Common Mistakes

This page lists the mistakes the skill is designed to catch before they reach a paper, rebuttal, thesis, or defense.

Fabricated References

Risk pattern:

Add five citations supporting this claim.

Better workflow:

1. Search for real papers.
2. Verify title, authors, venue, year, and claim.
3. Add only papers that were actually inspected.
4. Put unverified candidates in a pending list.

Inflated SOTA Claims

Risk pattern:

This method is state of the art.

Safer workflow:

• identify benchmark, metric, and protocol;
• compare against current baselines;
• check whether datasets and splits match;
• qualify the claim if evidence is incomplete.

Experiments Written as Completed Before They Exist

Risk pattern:

The experiment demonstrates a 12 percent improvement.

If the run has not happened, write:

The experiment is designed to test whether the method improves the target metric.

Invented Local Requirements

Risk pattern:

The school requires exactly five chapters and 30 references.

Only state this if the user provides the official template, advisor instruction, or department rule.

Rebuttal Overpromising

Risk pattern:

We will add all requested experiments.

Better workflow:

• separate feasible revisions from future work;
• promise only changes that can be completed;
• update the manuscript to match each rebuttal promise.

中文说明

这个页面列的是最需要提前拦住的学术写作错误。

常见风险：

错误	风险	正确做法
虚构引用	引用不存在或未验证论文。	先检索、核对标题作者 venue，再引用。
夸大 SOTA	没有完整 benchmark 和 baseline 支撑。	限定数据集、指标、协议和比较对象。
实验没做却写成完成	把 planned 或 expected result 伪装成 completed。	按 Planned、Preliminary、Completed 标注状态。
编造本地要求	没有官方模板或导师要求却写成学校规定。	只使用用户提供的真实要求。
Rebuttal 过度承诺	承诺无法完成的实验或修改。	区分可完成修改、解释、limitation 和 future work。

安全写法优先于漂亮但不真实的写法。证据不足时，应降低 claim 强度，而不是把语言写得更像论文。