Michal Valko : Collaborate

PhD and Postdoc Supervision

Over the years I have had the privilege of supervising many wonderful PhD students and postdocs at Inria and through collaborations with universities (MBZUAI, Ecole Polytechnique, ENS Paris-Saclay, Comenius University, U. Copenhagen). Some of the topics I am currently most excited about:

• Massive recursive self-improvement for language models
• Post-post-training: going beyond fine-tuning with test-time computation and inference-time search
• Reinforcement learning from human feedback and LLM alignment
• Game theory and multi-agent learning
• Bandit algorithms and sequential decision-making
• Self-supervised and representation learning

I want to be transparent: in recent years, I have only been able to take on new students in very specific circumstances — typically when there is already a strong personal connection, a well-defined research direction, and a clear co-supervision arrangement in place. My current commitments mean I simply cannot give a new student the attention they deserve unless the setup is right from the start.

If we have not had a chance to interact before, the most realistic path is to come with your own funding already secured (or a concrete plan for it) and, ideally, a local academic co-advisor at the institution where you will be based. I realize this is a lot to ask, and I genuinely wish I could be more flexible — but I have learned the hard way that taking on students without the right structure in place is unfair to everyone, especially the student. Co-supervision with other researchers is very welcome — many of my most rewarding supervisions have been joint efforts, and it often leads to richer research.

Have a look at current and past students to get a sense of the kind of work we do together.

Research Collaboration

I am always happy to explore research collaborations, co-supervision arrangements, and visits from fellow researchers. My recent work focuses on massive recursive self-improvement, post-post-training (test-time computation and inference-time search), LLM alignment, game-theoretic approaches to RLHF, reinforcement learning theory, and reasoning in large language models — see projects for what is currently on my mind.

If you see an interesting overlap with your own work, I would love to hear from you. A short note with a concrete idea or question is the best way to start a conversation. I particularly enjoy collaborations that bring together different perspectives — whether from mathematics, computer science, neuroscience, or industry applications.

Speaking and Keynotes

I enjoy giving talks and regularly speak at international conferences, summer schools, corporate events, and government forums. Recent examples include the Global AI Show, GLOBSEC Forum, Machines Can Think, RLSS, and various AI summits around the world. You can find the full list on the talks page.

If you would like me to speak at your event, please send me an email with the event name, date, location, expected audience, and a suggested topic. The more concrete and specific the invitation, the faster I can give you a clear answer. I try to be responsive and will let you know quickly if it is a good fit.

I should be upfront that my schedule is quite packed, and I have found that the smoothest collaborations happen when the organizational overhead on my side is minimal — ideally just showing up and giving the talk. I am unlikely to be able to handle reimbursement forms, coordinate logistics calls, or manage back-and-forth arrangements in a timely way, and I would not want to hold things up on your end. So the more your team can handle directly, the easier it will be for both of us to make this work.

On a practical note: I do not send slides in advance, do dry runs, or attend speaker briefings on how to deliver a talk. I like to decide the final shape of my presentation very close to the actual talk — sometimes just minutes before — so that the content is as fresh and relevant as possible. I am happy to share a live link to my slides that your team can access at any time, and this has always worked smoothly. All you need is a reliable internet connection in the venue. If that is not possible, I can always bring my own laptop. There have never been technical issues with this approach, and it means you get the best version of the talk rather than a stale draft sent weeks earlier.

One thing I feel strongly about: when you invite someone as a keynote speaker, that invitation should mean what it says. I have occasionally encountered situations where an organizer extends a keynote invitation and then, after the speaker has committed, follows up asking them to sponsor the event, cover their own travel, pay a registration fee, or fill out funding applications. I consider this a serious breach of trust and a practice that undermines the academic community. We all benefit when conferences are run in good faith — speakers give their time and expertise, and organizers provide the platform and the support to make it happen. If your event cannot cover basic speaker expenses, please be upfront about that from the very first message so we can have an honest conversation. What does not work is changing the terms after someone has already said yes.

Czech and Slovak AI Community

Supporting AI activities in Slovakia and the Czech Republic is something very close to my heart — I have been involved in various ways for over two decades, from organizing math seminars and mentoring students to speaking at local conferences, advising on national AI strategy, and helping build bridges between the CEE region and the global AI community. I am always happy to help where I can.

That said, the best way to make it happen is to treat it like any other international engagement: propose concrete dates, arrange the logistics, and lock things in early. I am almost never already in the region (“Nebudeš náhodou na Slovensku, v Košiciach a v Prahe? Nemáš náhodou cestu okolo?”) unless a trip has been specifically planned by the person who wants me to attend or speak at their event, so rather than checking whether I might happen to be nearby, it is much better to simply set up the visit from scratch. I have unfortunately had to cancel on wonderful people more times than I would like, usually because things were left tentative for too long and my calendar filled up. The more concrete and organized the invitation, the more likely I will be able to say yes — and I really do want to say yes.

Startup Advising and Investing

I am a Venture Partner at Sparkle Ventures, KAYA VC, and Formula VC, and an angel investor and advisor to AI startups across healthcare, materials science, procurement, and developer tools. See the full list on the service page.

If you are building something interesting with AI and think I could be helpful, I would be glad to hear from you. A short note about what you are working on and how I might contribute is all it takes to get started. I am especially drawn to teams tackling hard technical problems with real-world impact, and I enjoy working with founders who are curious, ambitious, and open to honest feedback.

Contact

The best way to reach me is by email — you can find my details on the homepage. I look forward to hearing from you.

Please note that I receive a high volume of email and am unable to respond to all messages in a timely manner.

• For media inquiries regarding our startup, please wait until the public announcement.
• For media inquiries regarding Inria activities, contact com-lille@inria.fr.
• For media inquiries regarding ENS Paris-Saclay activities, contact secretariat-mva@ens-paris-saclay.fr.

For speaking invitations, meeting requests, potential collaborations, reviewing requests, and other inquiries: if I have not responded within 1 week, please assume I am regretfully not available.

If we have an ongoing collaboration or conversation and I have not responded within 1 week, don’t hesitate to ping me again.