ARTICLE · 1083855
【AI&Agents笔记】AI Agent之间的互动方式:AI Agent 参与学术评审
AI Agent之间的互动方式：AI Agent 参与学术评审
Diverse Interactions Between AI Agents: AI Peer Review
Slides:
https://github.com/createmomo/Notes-on-AI-Agents
之前的文章：
从 OpenClaw 看 AI Agent 的运作原理 理解Context Engineering Context Engineering （LLM Summary & Observation Masking & Memory） 本可完成的任务在 Context Engineering 后失败了，怎么办？ 模型自主判断何时压缩Context 过滤Context和模型自主做Context Engineering AI Agent之间的互动方式
Previous Articles:
How AI Agents Work: The OpenClaw Case Context Engineering Explained A Core Technology Behind AI Agents （LLM Summary, Observation Masking & Memory） What to Do When Tasks an AI Agent Could Originally Complete Fail After Context Engineering Mechanisms for Deciding When to Compress Context Context Filtering and Autonomous Context Engineering Diverse Interactions Between AI Agents
关于这篇笔记的说明：
主要参考了Hung-yi Lee老师2026年的机器学习课程（https://www.youtube.com/watch?v=VqB8zMujdjM），其中一小部分内容根据个人理解可能有微小的调整； 保留了最核心的内容，所以比较简短，但也尽量保留了知识的完整。所以适合快速的回顾复习，当然也适合新手快速理解。
About This Note:
This note is primarily based on Prof. Hung-yi Lee’s 2026 Machine Learning course (https://www.youtube.com/watch?v=VqB8zMujdjM). A small portion of the content may include slight adaptations based on personal understanding; Only the core concepts are retained, making this note relatively concise while still preserving the completeness of the knowledge. Therefore, it is well-suited for quick review and revision, and also for beginners to gain a fast understanding.
目录/Contents:
AI参与审稿 (AI in Peer Review) 被接收论文的特点 (Characteristics of Accepted Papers)
1. AI参与审稿 (AI in Peer Review)
在前面我们讲了很多AI Agent和Agent之间是怎么互动的；同时，我们也在后面开始浅浅的探讨了人与AI Agent们之间的互动和关系。
In previous articles, we have explored many different ways in which AI agents interact with one another. More recently, we have also started looking at the interactions and relationships between humans and AI agents.
在这一篇中，我们会再看一看在评审论文的场景中，人与Agent又会发生什么样的故事。你可能会想问，AI审论文就是AI来评审论文呀，这和Agent之间的互动有什么关系呢？
In this article, we will look at another scenario: academic peer review. What kinds of interactions might take place between humans and AI agents during this process?
You might ask: if AI is reviewing a paper, isn’t it simply a matter of letting an AI evaluate the paper? What does that have to do with interactions between agents?
如果直接来看，好像似乎确实没有太直接的关系。但其实，在同行评审中，评审论文并给出意见和评分，这个环节只是同行评审中的一部分。从提交论文到论文是否被接受，中间还要经过综合参考不同审稿人或AI Agent的评审意见，甚至有时还有经过多轮考虑和斟酌。
At first glance, there may not seem to be much of a connection. In reality, however, reviewing a paper and providing comments and scores is only one part of the overall peer-review process.
From the initial submission of a paper to the final acceptance decision, the process may involve combining and comparing feedback from multiple reviewers or AI agents. In some cases, there may even be several rounds of discussion and deliberation before a final decision is made.
这过程中无疑是肯定有人介入的，但其中也涉及到了Agent的介入。
Humans are certainly still involved in this process, but AI agents are increasingly becoming part of it as well.
比如有的reviewer审稿者就是AI Agent,有的虽然生成说是人类审稿者，但他背后有没有借助Agent来帮忙呢？现在其实也很难说。除此之外，有的会议中虽然有AI和人类Reviewer并存，但是每轮审稿的决定，可能还是由人类来决定等等。
For example, some reviewers may themselves be AI agents. Others may officially be human reviewers, but it can be difficult to know whether they have used AI agents to assist them during the review process.
In review systems where both AI reviewers and human reviewers participate, humans may still be responsible for combining the different reviews, resolving disagreements, and making decisions at each stage of the process.
所以这整个过程虽然很难捋清楚Agent与Agent之间、人与Agent之间到底有怎样的互动、是怎样影响彼此之间的输出和观念的，但肯定在间接上Agent之间会以人为媒介，而影响彼此之间的输出，换句话说，Agent之间还是有可见或者“不可见”的互动的。（当然你可能对此看法会有不同或更好的解释，欢迎留言讨论啦~）
As a result, it becomes difficult to draw a clear boundary between agent-to-agent interaction and human-agent interaction, or to determine exactly how these different participants influence one another’s outputs and judgements.
What we can say, however, is that even when two agents do not communicate directly, they may still influence one another indirectly through humans acting as intermediaries.
In other words, interactions between agents may still exist, whether they are directly visible or effectively “invisible”.
Of course, this is just one way of interpreting what is happening here. If you have a different or better way of thinking about it, feel free to share your thoughts.
在25年就有这样一个会议，AI主要来写论文，并且AI也会介入审查投稿等。
In 2025, there was already a conference built around this idea: AI was not only heavily involved in writing the papers, but also participated in reviewing the submitted work.
从这个会议的并不结果来看，其实接受率并不高，还不到20%。
According to the published results of the conference, the acceptance rate was relatively low, at less than 20%.
从图中我们也可以看出，其中包括人类的评价和3个AI审稿者的意见。
As shown in the figure, the evaluation of each paper included both human assessment and reviews from three AI reviewers.
如果你对这些论文和哪些领域相关（比如AI领域中的哪些主题、其他领域）的数据有兴趣，可以看上面那张图，这里就不再赘述啦。
If you are interested in the research areas covered by these papers, for example, which AI topics they focused on or whether they extended into other disciplines, you can refer to the figure above. I will not go into those details here.
2. 被接收论文的特点（Characteristics of Accepted Papers）
那被接收的论文，他们有什么样的特点呢？需要注意的是，在上面的那个会议上虽然是有AI审稿者的介入，但最终做决定的还是由人类来做的，只不过人类会看一下AI的审稿意见。并不是说哪些论文被录取也是完全由AI来支配的。
What characteristics, then, did the accepted papers tend to have?
First, it is important to note that although AI reviewers were involved in the review process at the conference mentioned above, the final acceptance decisions were still made by humans. Human decision makers considered the feedback provided by AI reviewers, but the acceptance of a paper was not determined entirely by AI.
这些论文在被投稿前，作者是要明确说明，被投稿的这篇论文中每一部分（想实验假设、实验设计、数据分析、写作）到底有多少AI的介入。当然，如果一篇论文又能被接收，同时，人类的介入又很低，则说明AI确实可以在一些科研任务上面减轻人类的负担。
Before submitting a paper, authors were required to specify the extent of AI involvement in different parts of the research process, including the formulation of hypotheses, experimental design, data analysis and writing.
If a paper was ultimately accepted while involving relatively little human contribution, this would suggest, at least to some extent, that AI may be able to reduce the human workload in certain research tasks.
上面的图表中的下方我们可以看到，有4个档位，A,B,C,D，A档表示那部分的工作量有95%以上都是人工完成的，而D档则表示这部分的工作有95%以上都是AI完成的。
In the figure above, the level of human and AI involvement is divided into four categories: A, B, C and D. Category A means that more than 95% of the work in that part was completed by humans, while Category D means that more than 95% was completed by AI.
那结果到底如何呢？
What, then, do the results reveal?
从这个结果分析上大体可以看出，在那些被接受的论文中，人类的参与程度还是高于所有投稿的。除此之外，我们也可以很明显的发现一个规律，在论文的初期工作中（比如想一想论文的实验假设、点子的设计、实验应该怎么设计等），这些工作的中，人类的参与还是比较多的。这说明人类在科研的前期工作中，还是起到了非常非常重要的作用。
Looking at the overall results, the accepted papers generally showed a higher level of human involvement than the submissions as a whole. Another fairly clear pattern can also be observed. During the earlier stages of research, such as developing hypotheses, generating research ideas and designing experiments, human involvement tended to be higher.
This suggests that, at least according to these results, humans still play a very important role in the early stages of scientific research.
所以，虽然现在AI Agent确实很厉害，但投稿者自己的想法，或者我们不仅仅局限在科研这个场景，使用AI Agent的人，如果能有自己的想法和独特的点子，还是非常重要的。AI确实很厉害，但不一定一直能产生真正有用的点子。
Although AI agents have already become highly capable, the researchers' own ideas remain very important. The same principle can also be extended beyond academic research. Whatever task an AI agent is used for, it remains important for the user to form their own judgement and contribute original and valuable ideas. AI can be extremely capable, but this does not necessarily mean that it will always generate genuinely useful ideas on its own.
总之，人类的引导仍然是必不可少的。
Overall, at the current stage of development, human guidance remains an essential part of the process.
在今年26年，受到上面25年会议的启发，出现了这样一个会议，类似的，在这个会议的投稿中，AI承担着主要的作者、以及、审稿工作的角色们。
In 2026, another conference emerged, inspired by the event held the previous year.
Similarly, AI played a major role in both paper submission and peer review at this conference. It could serve as a primary contributor to the generation of papers and also take part in reviewing submitted work.
但不管怎么样，AI Agent再强大它目前也只是一个工具。到底它能做出什么样的事情和结果，又如何完成任务，目前我们作为使用者的人类的参与，仍然是很重要的一环。
However capable AI agents may become, they are still tools at present. The results they produce, and the way in which they complete a task, are still strongly influenced by how users define the goal, provide information and guide the process. Therefore, at least for now, human involvement remains an important part of the overall process.