
A new semester, and the first lecture is in the books for my class on the "Mathematical Foundations of AI Alignment". aaroth.github.io/cis-7000-ai… What does that mean? Good question. We have about a semester in which to figure it out.

CS prof at Penn. Researcher at Resolution. Author of The Ethical Algorithm (w/ Michael Kearns). I study ML, alignment, game theory, and uncertainty.

A new semester, and the first lecture is in the books for my class on the "Mathematical Foundations of AI Alignment". aaroth.github.io/cis-7000-ai… What does that mean? Good question. We have about a semester in which to figure it out.

I had Codex formalize our randomized calibrated multiaccuracy upper bound --- it now has a lean proof. Really very easy.
🚀 STOC 2027, a flagship conference in theoretical computer science, made all 1,665 submissions publicly available. So we launched the STOC 2027 Formalization Initiative on Prove2Me! Our AI agents have translated the main theorems and supporting lemmas of 1,496 submissions into Lean 4 statements! Some already have completed proof formalizations, with many more awaiting contributions. You don't need to know Lean or fully understand the underlying mathematics to get started! If you know how to work with AI agents, you can help formalize proofs. We invite authors, researchers, and the broader community to explore these formalized statements, check them against the original papers, and contribute proof formalizations! 📊 Explore the initiative: stoc2027formalization.vercel… 🤝 Contribute proof formalizations: prove2.me/collections/acm-st…


Had fun at the FOCS TCS+AI workshop yesterday. @littmath gave a spectacular keynote, followed by a great panel with @Knikct, Dan Spielman, @B1ar2n3a and @praveshkkothari. Big audience; no recording, but hopefully the event was useful to attendees. photo cred @thegautamkamath

Front row seats for a #FOCS2026 panel on the future of TCS&AI. After keynote by Daniel Litt (@littmath), he joins an all-star panel w Nikhil Srivastava (@Knikct), Dan Spielman, Barna Saha (@B1ar2n3a), Pravesh Kothari (@praveshkkothari). Organized & hosted by Aaron Roth (@Aaroth)

Here is the website for the TCS+AI Workshop: aaroth.github.io/tcs-ai-focs… Tell your friends!

If you are at New York tomorrow, come to TCS+AI workshop. The floor will be hot 😁.
Looking forward to tomorrow's TCS+AI Workshop at FOCS. @littmath will give a keynote, followed by a panel discussion with Dan Spielman, @littmath, @praveshkkothari, @B1ar2n3a, and Nikhil Srivastava. I'll moderate. Come with all of your questions, anxieties, and hopes!

Looking forward to tomorrow's TCS+AI Workshop at FOCS. @littmath will give a keynote, followed by a panel discussion with Dan Spielman, @littmath, @praveshkkothari, @B1ar2n3a, and Nikhil Srivastava. I'll moderate. Come with all of your questions, anxieties, and hopes!

🥁 Very happy to announce that ACM EC'27 will take place on July 12–16, 2027, at Northwestern University, Chicago. ▪️General Chair: Jason Hartline ▪️Program Committee Chairs: Annie Liang and Grant Schoenebeck Looking forward to seeing many of you there! @AcmSIGecom

Multicalibration is useful for lots of stuff - omniprediction, fast agreement, hard core set construction - but often overkilll. For these applications you only need global calibration, plus multiaccuracy. But does calibrated multiaccuracy actually require fewer samples to learn?

Woah, thats a lot of papers. Tired of the rat race? FORC that! THe first deadline is coming up, Nov 11: responsiblecomputing.org/for…
STOC 2027 received 1665 submissions. In a bold new experiment, all submissions (and author identities) are immediately made public. This is the most immediate and revealing posting of any conference I am aware of. Take a look and see what you find! stoc2027.sigact-ec.workers.d…

Here is the paper: arxiv.org/pdf/2611.03305. This is joint work with the excellent @LuJiuyao and Georgy Noarov.

In particular, randomization isn't needed for the closely related problem of multicalbation. But for calibrated multiaccuracy, there is a sharp polynomial separation between the minimax sample complexity when the model is constrained to be deterministic compared to when its not.

The answer is yes - at least if you don't have too many groups (otherwise no). But surprisingly the sample complexity depends on if you allow your model to randomize at test time! Thats kind of weird; most single distribution learning problems don't require randomization.


The only thing the Agreement Theorem uses about Bayesian rationality and priors is that the "belief" at each round of conversation is calibrated conditional on the one round prior. This you can achieve efficiently, without any distributional assumptions: arxiv.org/abs/2411.19791