Julia Kempe

MIT PhD student Shobhita Sundaram (@shobsund) & CDS Silver Professor Julia Kempe (@KempeLab), and a team from FAIR show how AI can generate its own practice problems to solve nearly impossible tasks. nyudatascience.medium.com/generating-the…

Our groups' ICLR presentations: 🎉Oral: OpenApps arxiv.org/abs/2511.20766 Soft Tokens, Hard Truths arxiv.org/abs/2509.19170 How reinforcement learning after next-token prediction facilitates learning arxiv.org/abs/2510.11495 From Concepts to Components arxiv.org/abs/2506.17052

Nice overview of recent self-distillation works out there!

The funny thing is not that GPT could give 3 simple, elegant proofs. It is that it knew what it means to prove "Fortnow-style"...

@polynoamial @HarvardMath Very apt quote from co-1stProver @scottnarmstrong

@HarvardMath AI isn’t replacing mathematicians today, but it is changing mathematics:

"The verdict, it seems, is in: artificial intelligence is not about to replace mathematicians. That is the immediate takeaway from the “First Proof” challenge—perhaps the most robust test yet of the ability of LLMs to perform mathematical research." scientificamerican.com/article/first-…

8/ Final note: thanks to the mathematicians behind #1stProof. Some fellow 1st-provers + colleagues: @AcerFur @OpenAI @merettm @DayShuai @littmath @nasqret @jleda_x @Tomodovodoo @c2v47 @BGrayzel @ArsSocraticaAI 1stproof.org #1stProof

7/ Recommendations (2/2): Build an audit registry. Machine-led literature search + AI-written papers = citation risk. We need a public “verification traces” layer (arXiv-adjacent): papers accumulate audit logs/certificates (models dissect/rewrite/check).

1/ #1stProof New write-up is live: “Takeaways from the First-Proof Trenches” (with @scottnarmstrong + @MunosRemi). Enjoy! kempejulia1.github.io/1stProof-Attem… Working across #1stProof we pulled together takeaways and recommendations for the scientific community.

Nikhil Srivastava Rachel Ward Shmuel Weinberger Lauren Williams Colleagues: @AcerFur @littmath @nasqret @c2v47 @jleda_x @davidbessis Comments and scrutiny are very welcome.

11/ Again, we thank the authors for creating such a fascinating testbed — and we genuinely appreciate careful checks and independent audits from the community. Mohammed Abouzaid Andrew J. Blumberg @MartinHairer Joe Kileel @TammyKolda @nick_sriv Paul D. Nelson Daniel Spielman

7/ *Third — interlude: “Humor from your bot.”* We found all bots prone to shortcuts and laziness. “Let’s wait until Feb 13 to see the proof” was one of the most frequently proposed options.

6/ *Second:* During literature searches, the models surfaced very recent references that appeared to prove key theorems they needed. Several turned out to be AI-generated papers. Lesson: literature search is acquiring a new level of difficulty in the age of AI slop.

5/ After working with our AI “scientist fleet,” we want to share a few takeaways. *First:* Even top frontier models produced many false proof manuscripts. Only careful back-and-forth auditing between models exposed the bugs. We hope we got them all!