Quantum Information Processing research in the age of AI

General
,
1

Dear colleagues,

Over the past weeks there have been more and more results in which AI models played a crucial role, not only in mathematics and theoretical computer science, but right at the heart of quantum information.

July 2026 alone saw a remarkable cluster of long-standing open problems settled: two-copy distillability of Werner states, open since about 2000, resolved by four independent papers in a single week (2607.21367, 2607.23416, 2607.24309, 2607.24479); bipartite bound information, a question Gisin and Wolf posed twenty-five years ago, answered days later; unclonable encryption, with two independent constructions, one unconditional, the other efficient, from Pauli eigenstates, appearing within days; and, in the same stretch, zero quantum and private capacity outside the PPT and antidegradable classes, sharp continuity of the conditional entropy, separability in polynomial time, a strong converse for stabilizer codes, and the Lee–Yang spectral gap (2607.10765, 2607.14401). I’m sure many more will appear in the coming days and weeks. Several of these were cracked by several groups within days of one another, illustrating a new capability that suddenly became broadly available.

We seem to be at the dawn of an “industrialisation of mathematics”, with a potentially huge impact on all of theory research, and our field is squarely in its path. None of us knows where this is heading, and that is precisely why now seems a very good time to discuss it together, as a wider quantum community, while we still have room to shape how we respond rather than simply reacting a year from now. For those interested, Robert Huang’s Simons talk “How To Respond To The Automation of Research” and Tobias Osborne’s recent post on proof inflation are good starting points.

Concrete questions that come to my mind:

  • what is a “publishable unit”, when a well-posed question can go to a complete, Lean-checkable proof over a weekend?
  • what is the role of QIP, TQC, QCrypt and other quantum conferences, and how should they treat primarily AI-generated results?
  • how do we organise reviewing, if program committees and the arXiv face a volume they cannot triage by hand?
  • what should authorship and disclosure look like, when a model cannot be an author but the human contribution may be a single prompt?
  • how do we keep access equitable, when frontier models are restricted by money and geography contraints, and open-weight alternatives are only now catching up?

But also more philosophical ones:

  • what is the human value of theory research in this new setting?
  • what is the human motivation for doing it?

Let’s discuss these fundamental questions together.

UPDATE (30 July 2026): I recommend everybody interested in these questions to join the ongoing discussion on Zulip at Sign up | Zulip

4 Likes
2

Hi @cschaffner Thank you for initiating this conversation! These are questions I have been pondering for some time, so here are my two cents:

Moving forward, AI will continually provide us with answers, limited only by the complexity of the problem and the model’s capabilities. However, I believe the core of human value will remain in posing original hypotheses and questions for AI to test or resolve. To formulate these novel questions, one will naturally need to build upon previous, AI-generated results. Therefore, leveraging AI for problem-solving does not threaten the fundamental human motivation for scientific discovery.

Addressing the most debatable aspect, i.e., the attribution of credit: imagine an open problem originally posed by Person A. Much later, Person B writes a prompt that leads AI Agent C to solve it. Following my previous logic, the credit should ideally be distributed in the following order: A > C > B. Accepting this hierarchy will require significant humility from Person B. One could certainly argue that crafting the prompt and guiding the AI is a creative and productive experience. However, as AI capabilities continue to advance, the significance of this intermediary role will likely diminish.

But above all else, ensuring universal access to and affordability of these AI models must be our primary imperative. Guaranteeing this access is the only way to maintain a truly level playing field in the future.

1 Like
3

Thanks for initiating this conversation! While we are looking at a time of upheaval across all of quantum information theory research, I am very optimistic that it will lead to something good.

I think it is fair to say that theory research as we know it is dead. It was good while it lasted, but we are looking at a future where every part of the scientific pipeline in quantum info will look different.

While the scale of this development is unprecedented, I think the fact that how we conduct theory research changes significantly with new tools is part of progress. At my alma mater, there were still volumes upon volumes of PRL abstracts in the physics library. Imagine doing research without the internet these days! Or to consider another example: in one of my research projects I ended up needing results by Slepian on the spectral concentration problem. These results were a tour de force, which included loads of manipulations of partial differential equations, special functions, and so on. This is a skillset that I for myself never developed because such questions are easily handled by Mathematica. From this point of view, LLMs are just a much more powerful tool.

Despite the fact that how we do it will completely change, I remain very optimistic about the science of the future. This is because my main motivation for doing theory research is to gain understanding. To this end, it doesn’t matter if a super nice argument is found by some genius, an LLM, or (I wish!) myself. The part of research that I would call “puzzling” is fun, the competition can also sometimes be fun, but that is not the core of it. At this point in time, I think humans can still distill underanding much better than LLMs. I read a blog post recently that talked about the very related task of canonization, where results are digested and presented in a coherent and fitting form. In quantum information, Mark Wilde is a master of this. He always goes through the motions of rederiving arguments he encounters in his own language, together with his books resulting in a sort of canonical library of arguments.

I think this point of view hints at some possible answers to the more concrete questions. I think our task as researchers is to find and convey this understanding. I expect, for example, that grant schemes that only involved a proposal and no talk will cease to exist and we might even see funders starting to run assessment centers to give out grants. When everybody can just have a great proposal written by AI, giving the money to the person demonstrating the best understanding seems a fair way to do it.

Consequently, conferences should also aim to convey understanding, no matter if it is based on human or AI work. I do not know how the actual format of this is going to look like, but maybe we will move away from technical abstracts to actual expositions of the intended talk with the goal of selecting talks that are interesting and stimulating.

As a general note: I think the senior people in the field, the people with the “clout”, have a duty to try their best in how they present AI-fueled research, because this will set the standard for the future. We do not necessarily need to agree beforehand, because when a reasonable person finds a reasonable way of doing this, people will automatically copy it and it will develop into a de-facto standard.

Or, simply put, let’s just wing it with the best intentions, and after some time we will have it sorted.

3 Likes
4
5

It seems accessing the Zulip discussion requires an invitation. Could you share a link which doesn’t require one?

Many thanks!

6

Yes, I’ve just updated it, I hope it works now!

1 Like
7

Hi, I’ve also tried to log in, but it still says that an invitation is required.

Any help would be much appreciated!

8

I like all the points discussed so I don’t think it necessary to repeat them but I did want to focus on your point on how we do things will have to change (and this also relates a bit to some of the discussion that appeared during Robert’s talk).

Indeed the information always “exists out there” and our goal as theory researchers is to understand it. But earlier, if the information wasn’t available in a textbook/in a paper/via a class, then we did the activity of research to discover/invent it. However now we will just use the AI to learn it, almost like attending a class and chatting to a teacher, which in my opinion is very different to the activity of doing research which I would argue includes what you mentioned in coming up with (part) of an argument yourself.

So I don’t really have many solutions, but I think even though it is important to preserve the culture of science, including the understanding, communicating and even applying science, I feel that the culture of theoretical research will have to invariably change (and can we even call it “research” at that stage?)

9

This seems key. The only motivation that really keeps the wheels of knowledge running.

This too. Speaking from personal experience. While I am currently only using AI for non-core tasks (more specifically, visualisation and user interaction), I sometimes find it hard to really know if something discussed in or even reflected upon after a conversation with AI is mine or not. Difficulty is bound to increase in collaborative settings.

&

Additional point. It would seem that the answer to many of the questions/issues in this thread change depending on whether one is thinking of prompted- or agentic-AI.

Agentic AI is a different beast in terms of motivations, ownership, and control.

It’s also likely more long-term relevant than prompted AI.

Ps. Hope this forum was open to non-theorists. Else, apologies for the intromission.

2 Likes