让每条建议都经得起追问：2026 丘奖季的 AI 辅助审阅Advice that can be checked: AI-assisted review in the 2026 Yau season
一篇论文补上了公式，前一轮的意见就应该改变。这个简单要求，恰好说明 AI 参与研究辅导时容易忽略的事：给建议不难，持续判断建议是否仍然成立，才需要认真核对。When a missing equation is added to a paper, the review should change. That simple expectation exposes a difficult part of AI-assisted research support: producing advice is easy; checking whether it still applies takes sustained attention.
今年丘奖季，YauLab Pro 相关工作涉及论文审阅、材料核对、修订对照和答辩准备。这篇文章从已核对的部分工作记录中提炼方法，不展示任何学生的完整个案。下文的情境均作一般化说明，不对应某个学生的课题、数据或评价。This Yau season, work associated with YauLab Pro has included paper review, submission checks, revision comparisons and defense preparation. This post draws methodological lessons from a limited set of reviewed records, without presenting any student’s case history. The situations below are generalized explanations, not accounts of an identifiable project, dataset or testimonial.
我们把这种工作方式称为 Evidence-first Copilot，证据优先的研究副驾驶。它从学生已有的材料出发，把意见连回证据，再由学生和老师核验。这个名称是对方法的概括，不是赛事认证。本文也不据此统计全年用户量、节省工时或参赛效果。We call this approach an Evidence-first Copilot: start with the student’s existing work, connect advice to evidence, and leave verification with the student and teacher. The name describes an approach, not a competition certification. This post does not establish annual user numbers, time savings or competition outcomes.
结论写得太满，具体满在哪里？When does a conclusion go beyond the result?
考虑一个通用的模型比较情境：某项误差指标变好了，另一个指标却没有同步改善。可以写清第一项的变化，但不能因此写成所有指标都更好。审阅要指出究竟是哪一句结论超出了哪一列数据，而不是笼统地说结论不严谨。Consider a general model comparison: one error metric improves while another does not. The first improvement can be reported, but it does not support a claim that every metric improved. A useful review identifies the sentence that goes beyond a particular result, rather than merely calling the conclusion weak.
比较之前，还要核对样本、模型版本和统计口径。带符号的平均偏差与偏差的绝对值不能混着读；没有相应检验，也不能把日常意义的明显改善写成统计意义的显著改善。Before comparing results, check the samples, model versions and statistical definitions. Signed mean error is not the same as its absolute magnitude. A descriptive improvement should not be presented as statistically significant without an appropriate test.
有些判断可以进一步落到算术上。设 Bias 为误差均值，SD 为按 n−1 作分母计算的样本标准差，n 大于 1。同一组、等权误差满足下面的恒等式。Some checks can be made directly through arithmetic. Let Bias be the mean error and SD the sample standard deviation calculated with denominator n−1, for n greater than 1. For the same equally weighted errors, the following identity holds.
RMSE² = Bias² + (n−1) × SD² / nRMSE² = Bias² + (n−1) × SD² / n
它来自误差平方和的分解。使用前须确认每项统计量来自同一批误差，且标准差口径一致。即使这项核对通过，也只能说明公布数字在这一层面自洽，不能代替原始数据和代码的复现。This follows by decomposing the sum of squared errors. All terms must use the same errors and the stated standard-deviation convention. Passing the check establishes a necessary arithmetic consistency, not a reproduction of the underlying experiment.
对应的修改动作也应有限而明确：保留能核对的结果，纠正过度概括，补齐数据与代码对应关系。不能为了让报告看起来更完整，替学生增加未经验证的结果。The revision should be specific and bounded: retain supported results, correct overstatements, and document how data and code connect. A more complete-looking report is no reason to invent additional findings for the student.
AI 的检查结果，也需要被检查AI checks need checking too
文本提取工具可能漏掉根号、上下标或作者信息。没有提取出来，不等于论文里没有。遇到这种疑点，审阅需要回到页面视觉层，确认它是识别遗漏，还是文件真的有问题。Text extraction can miss radicals, subscripts or author information. Missing from extracted text does not mean missing from the paper. A reviewer needs to inspect the rendered page to distinguish an extraction problem from a problem in the document.
若符号实际存在，就撤回误报。若最终 PDF 确实出现损坏符号、裁切标签或旧图号，就回到原始公式和绘图输出修复，再检查新导出的文件。两种情况不能用同一句缺项建议处理。If the symbol is present, withdraw the false alarm. If the final PDF really contains damaged symbols, clipped labels or outdated figure numbers, repair the source and inspect a fresh export. These are different situations and need different responses.
我们保留这一层复核，是因为自动检查提供的是线索。它不能自行证明原稿有错，更不能授权系统根据上下文猜补公式或改动研究数据。This verification step matters because automated checks provide leads. They do not, by themselves, establish an error in the original or justify guessing a missing equation or changing research data.
从分组名称，追到数据真正怎么走Follow the data, not just the split labels
一张表写训练结果，另一张表写测试结果，并不足以回答验证是否独立。审阅还需要问：原始记录是否重叠，参数选择是否用过测试信息，同一批采集产生的多个片段之间有什么关系？Separate tables labeled training and testing do not establish independent validation. Review also needs to ask whether source records overlap, whether test information influenced parameter choices, and how multiple segments from the same acquisition relate to one another.
可执行的建议，是请学生把原始记录、采集批次、切分区间、训练或调参用途，以及最终进入的结果表对应起来。每一步能追到实际文件，才有依据判断。证据没齐时，应写需要核实，不能直接指控数据泄漏。An actionable request is a mapping from source records and acquisition batches to split intervals, training or tuning use, and final result tables. Traceability makes judgment possible. Until evidence is available, an unresolved question should not become an accusation of leakage.
模型对照同样如此。用了相同数据，不代表实现、调参条件和比较口径已经一致。需要检查真实代码与运行记录；建议做某项核查，也必须和已经完成某项核查分开记录。Model comparisons need similar care. Shared data alone do not establish comparable implementations, tuning conditions or evaluation definitions. Actual code and run records matter, and a proposed check must be distinguished from one already completed.
研究迭代，不能靠重复发送旧报告Revision review means updating the judgment
把反馈放进研究工作流，意味着它会随材料变化。比如，原先缺少说明，后来补上说明，此时下一轮应核对说明是否有证据支持，而不应继续报告原来的缺项。这里讨论的是通用复核逻辑，不复述任何学生的修订经过。Integrating feedback into a research workflow means updating it as the work changes. If a missing explanation is added, the next review should examine its support rather than repeat the original missing-item finding. This is a general review principle, not a reconstruction of a student’s revision history.
至少要分清三个状态：还缺说明；说明已补但证据未核；相关证据已核实。笼统写一句已修改，会把这三种情况混在一起。At least three states need to remain distinct: an explanation is missing; it has been added but supporting evidence is unchecked; or that evidence has been verified. A single label such as revised loses these distinctions.
今年这部分工作的经验，推动我们把产品关注点放到意见与证据的对应关系上。审阅哪个版本、依据是什么、复核后判断为何改变，都比报告的页数更值得保留。这是产品方向，不代表所有项目都已完成整改或获得赛事认可。Lessons from this season’s work have focused our product thinking on the relationship between advice and evidence. Which version was reviewed, what supported the finding, and why the judgment changed matter more than report length. That direction does not imply that every project has completed revisions or received competition approval.
学生反馈与审阅反馈，要分开Student feedback is not the same as feedback to students
一份发给学生或家长的审阅意见，不是学生对产品的评价。准备好的答辩问题表，也不说明真实练习已经发生。修订稿出现变化，更不能单独证明学生满意或变化完全由 AI 带来。A review sent to a student or parent is not their evaluation of the product. A prepared defense-question sheet does not establish that a practice session took place. Changes in a draft alone do not demonstrate satisfaction or prove that AI caused them.
这篇公开总结不使用学生原话、聊天截图、个人反馈或满意度数字。我们保留的是反馈应怎样被处理的方法：问题说清楚，学生可以回应和质疑，复审根据新证据调整判断。没有经过核实的效果，不替学生作结论。This public reflection includes no student quotations, chat screenshots, personal feedback or satisfaction figures. It retains the method: explain the issue, allow the student to respond or challenge it, and revise the judgment when evidence changes. We do not infer outcomes on a student’s behalf.
AI 使用说明，要对应真实行为AI disclosures should match actual use
中国内地官网公布的 AI 使用规则要求事先获得指导教师许可，并如实说明工具及版本、使用环节与用途、时间及频率；使用 AI 时还要求提交相关聊天记录。研究主体仍须由学生本人完成，相关文献、公式、代码和结果需要人工核验。The AI-use rules published on the mainland China website require prior supervising-teacher permission, truthful disclosure of tools and versions, stages and purposes, timing and frequency, and relevant chat records when AI is used. Students remain responsible for the substantive research, with human verification of references, formulas, code and results.
丘成桐中学科学奖 AI 使用规则 / Official AI-use rules
核对重点应是实际使用记录、论文版本和贡献说明能否对应。检测分数不能代替这一过程，缺失的历史也不能靠生成对话补齐。The check is whether actual tool use, paper versions and contribution statements agree. A detection score cannot replace that process, and missing history must not be filled with generated conversations.
上面的具体说明对应中国内地官方页面，其他赛区须核对自己的当期要求。内地参赛规则还限制营利机构指导，不能因为服务被称为 AI 工具或审阅，就自行宣称获得豁免。Those specific requirements refer to the mainland China pages; other regions need their own current rules checked. Mainland rules also restrict guidance by for-profit organizations. Calling a service an AI tool or a review does not establish an exemption.
中国内地赛区参赛规则 / Mainland China entry rules
留下来的，是下一步能怎么核实What remains is a checkable next step
说数据不充分，就指出缺的是原始文件、划分记录还是实现证据；说结论过强，就说明哪项指标不支持它；材料发生变化，就重新判断。If data support is insufficient, identify the missing source files, split records or implementation evidence. If a conclusion overreaches, identify the result that fails to support it. When the material changes, review the judgment again.
这就是我们希望 AI 在研究工作流中承担的部分。学生保留研究与解释的责任，老师核对关键判断，工具帮助把问题指出来、把依据找回来。未解决的问题，不会因为报告写得完整就自动消失。That is the role we want AI to play in the research workflow. Students retain responsibility for their research and explanations, teachers check consequential judgments, and tools help locate issues and supporting evidence. An unresolved problem does not disappear when the report looks complete.