Very grateful to have received this Fellowship! Thank you to my mentors and collaborators for their support, especially Yee Whye Teh.

"Group Symmetry in PAC Learning" by Hayder openreview.net/pdf?id=HxeTEZJ… makes precise the intuition that learning with invariant hypotheses is equivalent to a reduced problem on a space of orbit representatives, which results in sample complexity benefits.

We have a really exciting line-up of pracs this year! 🚀 - bit.ly/3P397P6 We focused on providing more foundational/beginner-friendly content, while also introducing more diverse topics. 🧬🗺️🤖 In this 🧵, I will mention some of the amazing pracs lined up! 🔥

"Simplifying Transformer Blocks" ranks easily among my favorite research papers that I've read this year. Here, the authors look into how the standard transformer block, essential to LLMs, can be simplified without compromising convergence properties and downstream task

GNNs typically exploit permutation symmetry. Yet for learning tasks on a fixed graph, we show that enforcing active/approximate symmetries improves generalization. Check out our work "Approximately Equivariant Graph Networks" #NeurIPS23 (joint work w/ Ron Levie Soledad Villar )

# on shortification of "learning" There are a lot of videos on YouTube/TikTok etc. that give the appearance of education, but if you look closely they are really just entertainment. This is very convenient for everyone involved : the people watching enjoy thinking they are

Useful overview of attacks/defences in conversational LMs arxiv.org/abs/2402.09283

ASCII Art Jailbreaking of LLMs wild arxiv.org/abs/2402.11753

You will never be able to experience everything. So, please, do poetical justice to your soul and simply experience yourself. – Albert Camus

very nice video giving an overview of neural scaling laws youtube.com/watch?v=20P2jk…

Interesting, impressive and well written paper. Direct Ascent Synthesis: Revealing Hidden Generative Capabilities in Discriminative Models arxiv.org/pdf/2502.07753 Stanislav Fort Jonathan Whitaker

Well written and much needed paper on statistical analysis of AI evals. Rigorous results reporting should be more prevalent in the field. Long overdue. Adding Error Bars to Evals: A Statistical Approach to Language Model Evaluations arxiv.org/pdf/2411.00640 by Evan Miller