Nikhil Garg

A picture of Joe Halpern smiling in green shirt in front of a blue background.

Today arXiv remembers our colleague Joe Halpern, who was instrumental in founding arXiv's CS section.

Joe's passions ranged far & wide and we're lucky that arXiv was one of them. Joe, thank you for giving so much to arXiv - you are missed.

blog.arxiv.org/2026/02/27/remembering-joe-halpern

Care about preparing people to contribute responsibly to building the next generation of AI and technology?

Full-time (or at least 50%) lecturer position at Cornell Tech just posted, teaching computer science or related topics.

academicjobsonline.org/ajo/jobs/31698

Screenshot from the paper with a figure showing 15 scatterplots in a grid., evaluating LLM-as-judge on HELM. In each plot, one model is used as the judge. Each dot is another model; the y-axis is the accuracy inflation (compared to ground truth) of using the given model as the judge, and the x-axis is the model’s true accuracy. A vertical red line corresponds to the true accuracy of the judge. Each judge tends to inflate the accuracy of models that are less accurate than itself, especially models from the same provider or family

yeah! In this paper in ICML25, we found both directions of this -- in LLM as judge, using a bigger/more accurate model inflates accuracies bc of correlated errors, and using a worse model deflates them for the reason in the above paper

arxiv.org/abs/2506.07962

Wonderful, congrats!

Se Kenny's thread, and e.g., this blog post: transformer-circuits.pub/2022/toy_mod... by Anthropic. Kenny formalizes some of the claims made therein

Really glad to share this work by @kennypeng.bsky.social, toward better theoretical foundation for the linear representation hypothesis.

This call is still open. I am looking to recruit, as well as many other faculty at Cornell. We review folders as they come, and will send offers until all positions are filled.

Please share with your network 🙏

Thanks for sharing! Beautiful

This is an intriguing paper with numerous examples of using an LLM to generate formal proofs and ideas, and a nice cookbook for how to do it. Refreshingly, the paper stays away from hype and as far as possible states very clearly how the LLMs are used and the key nature of human-AI interaction.

What does that mean?

First Proof To assess the ability of current AI systems to correctly answer research-level mathematics questions, we share a set of ten math questions which have arisen naturally in the research process of the au...

Ten math problems with proofs known the authors. Proofs are encrypted until Feb 13. For all problems, authors claim both AI-based literature searches and zero-shot attempts at proofs failed. If you want to take a crack, you have until next Friday (2/13)!

Our paper “Inferring fine-grained migration patterns across the United States” is now out in @natcomms.nature.com! We released a new, highly granular migration dataset. 1/9

yeah, I'm not sure it's "useful yet" but perhaps "I should start learning to be on the cutting edge for when it is imminently useful"

Until seeing this paper, my view was, "Lean is promising but doesn't have the libraries for the type of math I typically do." And it seems like folks have made substantial progress in fixing that issue

I have been waiting for something like this to onboard onto Lean (alongside an LLM agent), and it seems like that moment is here

🚨New WP "@Grok is this true?"
We analyze 1.6M factcheck requests on X (grok & Perplexity)
📌Usage is polarized, Grok users more likely to be Reps
📌BUT Rep posts rated as false more often—even by Grok
📌Bot agreement with factchecks is OK but not great; APIs match fact-checkers
osf.io/preprints/ps...

ha I had seen snow before, but before I moved to NYC the coldest/densest place I'd really lived was SF

Wow! I admit I thought that bodies of water freezing over was a movies thing...

A search for factors for algorithm understanding results in multiple terms displayed as documents, including available, compact, and aligned. These are shown to be necessary and sufficient. Other, similar terms are shown in the background faded, like intuitive, rule-based, grounded, modular, linear, decomposable, accurate, symbolic, causal, and personalized.

Is the only way we can create algorithms that people understand to make them trivially simple? We argue, no.

People can predict the behavior of algorithms that are arbitrarily complex, if and only if they are available, compact and aligned.

arxiv.org/abs/2601.18966

Fairness in PCA-Based Recommenders

🎙️ I had a great time joining the Data Skeptic podcast to talk about my work on recommender systems

If you're interested in embeddings, aligning group preferences, or music recommendations, check out the episode below 👇

open.spotify.com/episode/6IsP...

There is a third-party app for this, you can find it here: bsky-follow-finder.theo.io

My impression with bluesky is that a lot of things exist but there is just no way of finding them.

@nkgarg.bsky.social

ah amazing thanks for sharing!

I've been playing with various side projects with codex. And even though I do not have time for next few weeks, this post is very close to sniping me into finally building this...

Yeah, I don't understand the need to ask certain questions during a talk

This is indeed more or less exactly what happened...

Check out our new paper at #AAAI 2026! I’ll be presenting in Singapore at Saturday’s poster session (12–2pm). This is joint work with @shuvoms.bsky.social, @bergerlab.bsky.social, @emmapierson.bsky.social, and @nkgarg.bsky.social. 1/9

Diversity is definitely top of mind for us!

Congrats team!

Thank you!

Yep we filtered those out!

bsky.app/profile/sjgr...