There are a lot of things which annoy me about ARR/ACL papers. Mostly content-level issues (such as poor evaluations and Hovy’s “LLM popcorn”), which I have written about elsewhere. But I am also unhappy with writing and presentation, especially when I see authors violate ARR rules and guidelines (see CFP and formatting guidelines). I really wish that authors followed the rules, and that reviewers rejected papers which did not follow the rules! I describe some of these problems below.

Readable font size in diagrams

As I have mentioned in a previous blog, text in diagrams must be readable. The ACL formatting guidelines state that

For any text or numbers in tables and figures, whenever possible, please use the font size of the document text. As a rule of thumb, any text or numbers should be clearly readable when the paper is printed on A4 paper. Submissions that abuse the font size or spacing for figures/tables may be desk rejected.

This rule is widely violated. I remember on one occasion a student showed me a draft paper which included a figure that needed 400% zoom to be legible. The student did not see anything wrong with this, but it is not acceptable to readers.

When I review a paper, I recommend desk rejection if the paper has unreadable figures.

Key information in appendices

Appendices in ARR papers have exploded in size, I often see 10-20 pages of appendices in an ARR paper. What is even worse is that authors put key material (eg literature review or important results) into appendices. Appendices are for supplementary material, especially material which supports replication. The ARR CFP states that

For example, pre-processing decisions, model parameters, feature templates, lengthy proofs or derivations, pseudocode, sample system inputs/outputs, and other details that are necessary for the exact replication of the work described in the paper can be put into appendices.

However, reviewers are not required to consider material in appendices. If the pseudo-code, or derivations, or model specifications are an important part of the contribution, or if they are important for the reviewers to assess the technical correctness of the work, they should be a part of the main paper rather than appendices.

When I review a paper, I follow the above rules and do not look at appendices. So if the literature review or key results are in an appendix, I automatically reject the paper.

I like TACL’s rules for appendices (5 pages max for details which support replication; 3 pages max for additional tables; no free text). I wish ARR would adopt something similar. This would help authors as well as reviewers, since many authors complain to me that some reviewers expect very large and detailed appendices.

Incorrect content from LLMs

I increasingly see papers and submissions (sometimes from respected researchers at major research universities) which contains hallucinated content which is incorrect, such as non-existent references. In one case, one of the authors told me that this had nothing to do with her, the relevant section had been written by a PhD student. The ARR CFP states that

In all cases, all authors are fully responsible for the correctness of their methods, results, and writing. They should check for potential plagiarism, both of text and code

If you add your name to a paper written by a PhD student, you are responsible for the paper being correct (and ethical), including LLM-generated content! If you have concerns (or have not even read the paper in question), you should not be an author.

Needless to say, if I review a paper with hallucinated content, I will reject it. In fields such as medicine, fraudulent content can lead to disciplinary action. In fact, the person who contacted me above worked for a university which had strict rules about this kind of thing. She could have been in a lot of trouble if someone reported this to her university (I did not do this).

Reference published papers and use DOIs

Many researchers, especially students, seem to prefer to read and cite Arxiv versions of a paper, even when the paper has been published in a peer-reviewed venue. I strongly believe that this is bad practice, since published versions often include improvements required by peer reviewers (blog).

I am not aware of an ARR policy about this, but I have seen other AI venues request that citations be to published papers if these exist. For example, the author guidelines for the computer vision conference CVPR 2026 states that

To reduce confusion, whenever citing papers that initially appeared on arXiv, the authors should check whether those papers had subsequently been published in a peer-reviewed venue, and to cite those versions accordingly.

Another issue is the use of DOIs for cited papers. DOIs are becoming the standard online reference for academic papers (including Arxiv papers) in most scientific fields. I think they work well, and the ARR formatting guidelines ask people to include DOIs for cited papers

All references are required to contain DOIs of all cited works when possible, or, as a second resort, links to ACL Anthology pages.

Most ARR papers seem to include DOIs for some references but not others, probably because some sources of references (such as ACL Anthology) include DOIs but others (such as Google Scholar) may not. I do not reject papers that lack DOI, but I do encourage authors to include them, this is good pratice.

Final thoughts

One thing that worries me is that a lot of the above is driven by research culture and expectations. Authors see nothing wrong with unreadable figures, literature reviews in appendices, etc, since they read a lot of other papers that do this. Of course poor research culture impacts content as well as writing.

It is hard to change research culture (blog), but publication venues can do this if they set clear rules and enforce them. ARR has established good rules and guidelines (although of course they could be improved, eg for appendices), lets insist that they be followed!