Pengzhou Wu ✈️ NeurIPS

Postdoc ＠sshimizu2006. Prev: Postdoc @ml_tuberlin, PhD in Statistics advised by Prof. @KFukumizu. #Causality/#Causalinference and Generative models. Zh: 吴鹏洲

[Pengzhou Wu ✈️ NeurIPS @PengzhouW]

#neurips2026 Accepted! https://t.co/AdH4z9qLci. This is the first identifiability result via concrete analysis of model symmetry of a nonlinear (and unsupervised) class, the first to admit multiple latent codes for every observation, and the first to allow discontinuous decoders.

[Pengzhou Wu ✈️ NeurIPS @PengzhouW]

@Anthony_Bonato the right *looks* cleaner/better and I think it should be easier to read too... but, no, I found the left is in fact easier to *read*, the right is too "blurred" when I really read it.

[Pengzhou Wu ✈️ NeurIPS @PengzhouW]

Hmm, seems the ICLR assignment respects neither my bidding nor the initial matching. I am fine with the assigned papers tho.

[Pengzhou Wu ✈️ NeurIPS @PengzhouW]

I support this stance whole heartedly, in fact on any AIGC. This is much more productive.

[Peter Richtarik @peter_richtarik]

@SincDavidson First, do not believe any AI detection tool. Second, do not worry whether AI was involved or not. Judge the paper, judge the science. Basically, do not discriminate on the basis of color, sex, religion, and whether the author is or is not a human (or any combination thereof).

[Peter Richtarik @peter_richtarik]

@SincDavidson First, do not believe any AI detection tool. Second, do not worry whether AI was involved or not. Judge the paper, judge the science. Basically, do not discriminate on the basis of color, sex, religion, and whether the author is or is not a human (or any combination thereof).

[Pengzhou Wu ✈️ NeurIPS @PengzhouW]

@insou This seems true for every senior researcher in universities. Maybe one day I will be something like an independent researcher to avoid non-research work. I am not sure and not sure how tho...

[Pengzhou Wu ✈️ NeurIPS @PengzhouW]

@StatsPapers 😁https://t.co/WRyb47AtCV

[Pengzhou Wu ✈️ NeurIPS @PengzhouW]

#neurips2026 Accepted! https://t.co/AdH4z9qLci. This is the first identifiability result via concrete analysis of model symmetry of a nonlinear (and unsupervised) class, the first to admit multiple latent codes for every observation, and the first to allow discontinuous decoders.

[Pengzhou Wu ✈️ NeurIPS @PengzhouW]

@torgeirlysen Yup, esp GPT. They can nitpicking fixable details. Seems easy to add some sys prompts to address this tendency.

[Pengzhou Wu ✈️ NeurIPS @PengzhouW]

Think a solution to the flood of LLM-written papers is not to ban them, but to allow (pure!) LLM reviews. Between a mediocre human reviewer and a good LLM (GPT 5.4 or Opus 4.5 above), I'd choose the LLM to review my draft w/o hesitation!

[Pengzhou Wu ✈️ NeurIPS @PengzhouW]

@ThomasWolgast "Worst case is a human blindly using an LLM to create a review." Yup, and we know this is what is happening more than 80% (?) cases nowadays. Yes, I think some "supervision" given to both LLM and human reviewers will be very useful.

[Pengzhou Wu ✈️ NeurIPS @PengzhouW]

@ThomasWolgast I completely agree. By "pure" LLM reviews, I mean it's transparent to all humans involved (the journal, the possible human reviewers, the authors, the readers). I think human reviewers can join after a paper is not rejected by an LLM concensus.

[Pengzhou Wu ✈️ NeurIPS @PengzhouW]

I meant, even if my paper is fully human written, tho I won't do so nowadays.

[François Fleuret @francoisfleuret]

You will have to spend a year isolated, everything will be provided, food, etc. but no contact with the external world.

You have to choose your companion between

[Pengzhou Wu ✈️ NeurIPS @PengzhouW]

@sht_takahashi おめでとうございます。こっちもです！（今時の推薦アルゴリズムはすごい

[Pengzhou Wu ✈️ NeurIPS @PengzhouW]

Yes, purely LLM-written papers come out like a flood (of rubbish?), but they are easy to recognise by humans anyway.

[Pengzhou Wu ✈️ NeurIPS @PengzhouW]

LLM-assisted writing is good for review system in one aspect at least. Dressing-up and name-dropping used to be a skill, but now everyone's LLM does it, so it's worthless. Reviewers hopefully will be immune and focus on correctness and honest motivation.

[Pengzhou Wu ✈️ NeurIPS @PengzhouW]

Why do I find so many errors? Diligence: for example, I spent 9 hours on one paper, the topic of which I am not very familiar.

[Pengzhou Wu ✈️ NeurIPS @PengzhouW]

I've rejected more than 5 conference papers in a row, and my bar isn't high: I only reject when I can see clear errors or serious gaps. Significance/impact rarely entered consideration. #NeurIPS2026 had the highest reviewer agreement I've ever seen. I credit AI's error-hunting.

[Pengzhou Wu ✈️ NeurIPS @PengzhouW]

I joined #NeurIPS2026 experiment yet got zero AI-assisted papers. Seems impossible under true random assignment. Maybe an unstated constaint that experienced / past-best reviewers go pure-human?

[Pengzhou Wu ✈️ NeurIPS @PengzhouW]

Oh this time in my review batch, after rebuttal, there is a paper passed the correctness bar, and it got accepted. My own paper has large score variations tho, but because it's purely theoretical.

[Pengzhou Wu ✈️ NeurIPS @PengzhouW]

Now the reviewer could let AI nitpick instead. But if the paper has no fatal flaw, AI surfaces only "real, but minor or rebuttal-addressable" issues.

[Pengzhou Wu ✈️ NeurIPS @PengzhouW]

I think #neurips2026 AI-assisted review experiment is a success. One signal: it accepted more marginal-but-correct papers. For good or bad? I tend to the former, tho not sure.

[Pengzhou Wu ✈️ NeurIPS @PengzhouW]

Flip side, as an author: AI helps reviewers actually understand difficult papers (mine, for one), and it defuses bias. The most frustrating case reviously: a hostile reviewer could hide behind vague or absurd reasons, making rebuttal nearly impossible.