Grigoris Velegkas

Research Scientist at Google; prev. Computer Science PhD student at Yale University:

[Grigoris Velegkas @gvelegkas]

Excited to share some of the TCS results we’ve obtained recently using the “Cogentic” harness on top of Gemini 3.1 Pro and some early checkpoints of Argon. We focused on areas of our expertise, like algorithmic game theory and learning theory. See Aranyak’s post for details!

[Aranyak Mehta @aranyak_mehta]

These are exciting times for AI-Math and AI-TCS. In our team at @GoogleResearch, @GoogleDeepMind we have been working on a multi-agent orchestrator for theorem proving we call "Cogentic". We built it inspired by our own experience on how we conduct research: dividing work among agents such as diverse provers, adversarial verifiers, strategy overseers, literature reviewers, and paper writers.

We're happy to share the paper describing the harness and some of the open problems we've solved using it. A typical run is quite efficient, making O(100) to O(1000) calls to Gemini (using 3.1 Pro and an earlier version of 4 Argon). We can start with just the problem statement without any hints, and the system can autonomously output human-readable proofs.

https://t.co/QY3dZwdWDo

Results next:

(1/2)

[Grigoris Velegkas @gvelegkas]

@roydanroy @shai_s_shwartz @amit_daniely This was for distribution learning. For multiclass classification, this paper https://t.co/QiV4qid6Uh had shown that the DS dimension characterizes learnability and Chirag showed the tight dependence on DS (and overall tight complexity)

[Shai Shalev-Shwartz @shai_s_shwartz]

Huge congrats to Chirag Pabbaraju on this beautiful result!
The new paper shows the DS dimension fully characterizes multiclass learnability - a true analogue of VC dimension beyond the binary world.

Thrilled (and a bit surreal) to be the “S” in DS :)
link 👇
@amit_daniely

[Stat.ML Papers @StatMLPapers]

The Geometry of Efficient Nonconvex Sampling https://t.co/Uthg2hY7Ef

[Felix Zhou @felix_zhou_cfz]

Previous densest subgraph algorithms in the continual release model incur O(log n) error and space overhead compared to static counterparts.
By densifying the input graph, this overhead can be removed!

To appear in SOSA'26.

[Amin Karbasi @aminkarbasi]

My brilliant student @AnayMehrotra and amazing post-doc Alkis Kalavasis just won the Best Paper Award at COLT 2025! Clearly, they’ve cracked the code: do great math, write a killer paper, take home the trophy.

Huge congratulations to both of them.

Now please don’t forget your humble advisor when you’re famous!
https://t.co/NRuGgpdJSB
Paper by: Yang Cai, Alkis Kalavasis, Katerina Mamali, Anay Mehrotra, and Manolis Zampetakis.

[Anay Mehrotra @AnayMehrotra]

We are organizing tutorial on Language Generation at #COLT 2025!

Visit our website (link below) for references and materials; content is updated regularly, so check back for the latest.

Organizers: Moses Charikar, Me, Chirag Pabbaraju, @_cpeale, @gvelegkas

See you in Lyon!

[Charlotte Peale @_cpeale]

Come join us tomorrow at #COLT2025! 📍 Our tutorial on Language Generation covers the recently proposed "generation in the limit" framework of Kleinberg and Mullainathan and the exciting space of recent work building on it.

⏰ 9:30am - 12:00pm, Room C
🔗 https://t.co/OzpQ8Fv6Nf

[Stat.ML Papers @StatMLPapers]

(Im)possibility of Automated Hallucination Detection in Large Language Models https://t.co/dN64FAstaU

[Kasper Green Larsen @kasperglarsen]

Significantly improved sample complexity of replicable boosting via a majority of smooth-boosters.

New arxiv preprint: https://t.co/OKuMmRTd12

[Anay Mehrotra @AnayMehrotra]

We want language models that do not hallucinate

We want language models that have breadth (i.e., no mode-collapse)

Jon Kleinberg-@m_sendhil asked: Can we get both?

Alkis, @gvelegkas, and I show this is impossible: https://t.co/m28754LyHT

🧵(1/3)