This morning, I overheard two two theoretical computer scientists saying: "but ... constants do matter". Yes, they do :)

نوروز مبارک!

Now that Talagrand has won the Abel Prize, once again I read this biography of him. One aspect that I found very funny is Talagrand's description of Parisi using the replica method to analyze the Sherrington-Kirkpatrick Model :) michel.talagrand.net/longbio.pdf

Looking to keep up-to-date with the latest LLM jailbreaking attacks and defenses? Look no further! Introducing JailbreakBench -- a standardized, reproducible benchmark for jailbreaking LLMs.

The best middle east history podcast I have seen so far: youtube.com/@Unravellingth…

This paper is accepted to #ICML2024!

Behrad Moniri, Hamed Hassani: Signal-Plus-Noise Decomposition of Nonlinear Spiked Random Matrix Models arxiv.org/abs/2405.18274 arxiv.org/pdf/2405.18274

Evaluating the Performance of Large Language Models via Debates arxiv.org/abs/2406.11044

Conditional validity and length (set size) efficiency are crucial aspects of conformal prediction. We developed CPL, a principled framework that optimizes these two objectives simultaneously. 📄 Paper: arxiv.org/pdf/2406.18814 🖥️ Step-by-Step in Python: tinyurl.com/mr24d3cc

We study a nonlinear spiked random matrix model where a nonlinear function is applied elementwise to a noise matrix perturbed by a rank-one signal. We establish a signal-plus-noise decomposition for this model and use it for problems in signal recovery and stochastic block models

We propose an automated benchmarking framework for LLMs where the evaluation of models is based on debates between models, judged by another LLM. This method gives rankings that align well with popular rankings based on human input, eliminating the need for human crowdsourcing.