Post

The Call for Papers for #STOC2027 is up! Importantly, the PC "will place substantial weight on the quality of exposition, and clarity of technical arguments and proofs" Also: - Public posting requirement - Required video submission and more. Deadline: ⏰ Nov 2, AoE acm-stoc.org/stoc2027/sto...

I like most of them but not sure how I feel about reviewers being allowed to use LLMs

I agree — not something I am personally happy about. But decisions were made collectively, after thorough discussions.

Ah at least author consent is required

My bad: authors are required to consent to use of LLMs to review. Quote “Authors must explicitly consent to both uses as part of the submission process.”

Yeah...

WAIT WHAT. I read that statement very different. I read it as "authors can choose to consent or not, and must make that explicit." Not "a condition of submitting is that you consent to LLM reviews." Are you telling me the intention is the latter??

Yes it appears so. I had the same misreading…

Confirmed with the PC chair that it is indeed that authors must consent as a precondition of submitting.

@ccanonne.github.io - I might as well say this in public now. AI-assisted reviews are not going to have a good start; my worry is that people will review things without making an attempt to struggle with the paper.

I don't disagree with this (or, rather, I totally agree, and hope to be proven wrong). As I said, the current policy was discussed and voted by the PC members: I may not agree with all, but that's the majority decision.

I was wondering if this is connected with requiring the paper is public (on the arXiv or equivalent) at the time of submission. By which I mean: if it’s on the arXiv then the LLM scrapers have access anyway, so feeding the paper into an LLM for a referee report comes with minimal additional damage.

But what you’re providing as an expert feeding a paper to an LLM is far far more valuable. You are providing YOUR TRAIN OF THOUGHT. As a referee you will engage in dialogue with the LLM about the article. This is immensely valuable future training data for the AI companies.