Chat screenshots reveal the hidden rules of AI review! AAAI 3000 yuan is strong accept?

Just as the AAAI 2023 paper submission deadline was approaching, a screenshot of an anonymous chat in the AI ​​submission group suddenly appeared on Zhihu.

One of them claimed that he could provide "a strong accept" service for 3,000 yuan.

As soon as the news came out, it immediately aroused public outrage among netizens.

However, don’t worry yet.

Zhihu boss "Fine Tuning" said that this is most likely just a "buzz".

According to "Fine Tuning", greetings and gang crimes are unavoidable problems in any field. With the rise of openreview, the various shortcomings of cmt have become more and more clear. The space left for small circles to operate will become smaller in the future, but there will always be room. Because this is a personal problem, not a problem with the submission system and mechanism. Introducing open review and bid random matching may be helpful.

And his personal view is that it doesn’t matter whether the article is published in NeurIPS, AAAI or IJCNN. Just publicize it enough to let everyone know what you are doing. The citation will prove whether this thing is useful. .

So, is this trend prevalent in today’s conference paper review? And has anyone gained "fame and fortune" from this?

Everything can make you rich

Before discussing, let’s briefly talk about what this bid is.

In order to facilitate the assignment of reviewers, some conferences will set up a bidding session to allow reviewers to select articles of interest. The system can then allocate papers based on the selection of these reviewers. However, reviewers will only see the paper title and abstract at this stage.

And there is a lot to be said about this.

Various anonymous revelations

Another anonymous netizen on Zhihu said , the layout is too small, as evidenced by WeChat chat screenshots, in fact, even the best papers can be bid by each other, this trivial matter is nothing.

Some netizens said that when submitting to AAAI this year, they were asked to fill in conflicting users, conflicting domain, etc. in a serious manner, but in fact it was of little use.

This anonymous Zhihu user even proposed an amazing solution: using graph neural networks, recommendation systems, etc. to automatically mine previous With the relationship network centered on each tutor, and then checking the bid data, review data, and scoring status of major conferences, abnormal nodes and relationships can be easily dug out, because they shine like stars in the graph neural network.

Or simply hold a top bid data mining competition.

However, some netizens pointed out that this so-called "3,000 yuan package SA" is actually buggy:

You must know that even if you forcefully recommend it, other reviewers will give it to them for free if they add a borderline and a weak reject. After all, this thing is not based on average admission scores. And if the differences are huge, AC will not be able to give it.

So even if you have to give money, you must use it on the "blade". If you think about it, then you have to find SPC or AC. But at this time, you need to consider two questions (actually one): First, do you think you can get someone of this level for just 3,000 yuan? Second, if someone really does it for you, will they only charge you 3,000?

Of course, in addition to complaining, there is also a senior reviewer who is "brave in his righteousness" and said that if he sees a reviewer giving "hydrology" a high score, he will directly strongly reject it. , and then write a short composition and diss it fiercely, and now the template has been drawn up.

#It is said that all conferences and journals have similar problems, but in different proportions and severity. It is inconvenient to say more in this article.

The suggestion of an anonymous user is true and talented

The review is random How random is the AI ​​review process?

In 2021, NeurIPS conducted a review consistency experiment.

According to the video introduction of "Electric Phantom Alchemy" by the Hong Kong Chinese CS doctor, in this experiment, there is a 10% probability that the same manuscript will be sent to two groups of reviewers. The reviewers in the group are unaware of each other's existence.

Two groups of reviewers will score separately, and two field chairs will make a decision on the scoring results. In this way, we can quantitatively analyze how likely it is that the same manuscript will receive different review results.

The results show that the probability of acceptance by the first group of reviewers and rejection by the second group of reviewers is 52%, which is more than half.

And if the first group of reviewers rejects, the probability of the second group of reviewers also rejecting is 83%.

This result is also consistent with Li Feifei's point of view.

Li Feifei once published a very famous article discussing review, and put forward these two points of view: 1. The review process is highly random. 2. Poor papers will get poor review comments. This is a golden rule that does not change with time and randomness.

NeurIPS officials also conducted such a comparison. Each paper was not reviewed and decided randomly, and then made based on the acceptance rate and disagreement rate. A curve, the horizontal axis is the acceptance rate, and the vertical axis is the disagreement rate.

As can be seen in the figure, if the reception rate is very low or very high, the divergence rate is not large.

The difference rate in the middle is the largest. If the acceptance rate is 50%, the difference is the largest.

Both AAAI and IJCAI adopt this model, using a high rejection rate in exchange for smaller randomness.

Finally, "Electric Phantom Alchemy" gives such a summary of the conclusions about artificial intelligence review.

AAAI on the cusp of the storm

At the beginning, the review results of AAAI 2022 were controversial.

Someone said on social media that his paper got 4 accepts, but was finally rejected.

The author below stated that the four reviewers all gave opinions of "acceptance bias" or "acceptance", and all of their review comments were positive. : The idea is interesting, the model is solid enough, and the experiments and analysis are comprehensive enough.

But the final result was: rejection.

For AAAI 2023, the two-stage double-blind review format is still used.

In Phase 1, each paper will receive two reviewers. If both opinions received are negative, the paper will be rejected outright. The remainder will proceed to Phase 2.

In Phase 2, each paper will receive at least two new reviewers. New reviewers will not see the Phase 1 review results until they submit their comments.

The authors of papers entering the second stage can respond to all review comments. Discussions between the Program Committee and the Senior Program Committee will be based on feedback from the authors.

As for the specific operation method, you can refer to the following paper.

This paper proposes a novel reviewer-paper matching method and has been applied (in whole or in part) at AAAI 2021, ICML 2022, AAAI 2022 and IJCAI 2022 ).

##Paper address: https://arxiv.org/abs/2202.12273In the end I can only At present, it seems that there is not much work that can be done when submitting a paper. The only thing a contributor can do is to try his best to write a good article and make his paper become the rightmost part of the NeurIPS experimental curve.

The above is the detailed content of Chat screenshots reveal the hidden rules of AI review! AAAI 3000 yuan is strong accept?. For more information, please follow other related articles on the PHP Chinese website!