
Meet the first NUS Computing Ascent Fellows: nine early-career researchers from around the world, visiting us in Singapore this January: comp.nus.edu.sg/bytes/nine-e… #NUSComputing #AscentPrize #2027AscentFellows

Professor of Computer Science at UCLA @UCLAComSci; working on Artificial Intelligence

Meet the first NUS Computing Ascent Fellows: nine early-career researchers from around the world, visiting us in Singapore this January: comp.nus.edu.sg/bytes/nine-e… #NUSComputing #AscentPrize #2027AscentFellows

Make sure to check out our paper!
📄 CoRL 2026 Spotlight: Meta-Ctrl predicts robot plan failures before the LLM task planner writes them. Exact inference, no sampling. Llama 3 8B beats GPT-5.5 & Claude Opus 4.8 on VirtualHome. 🎤 Talk 8:30 am · 📍 Poster #104, 10:30–12 🧵

Come check out our spotlight and poster talk today! 🎊
📄 CoRL 2026 Spotlight: Meta-Ctrl predicts robot plan failures before the LLM task planner writes them. Exact inference, no sampling. Llama 3 8B beats GPT-5.5 & Claude Opus 4.8 on VirtualHome. 🎤 Talk 8:30 am · 📍 Poster #104, 10:30–12 🧵

📄 CoRL 2026 Spotlight: Meta-Ctrl predicts robot plan failures before the LLM task planner writes them. Exact inference, no sampling. Llama 3 8B beats GPT-5.5 & Claude Opus 4.8 on VirtualHome. 🎤 Talk 8:30 am · 📍 Poster #104, 10:30–12 🧵

UCLA Computer Science is Hiring! The UCLA Computer Science Department at the Samueli School of Engineering invites applications for multiple tenure-track Assistant Professor positions in all areas of computer science. For more information and to apply: recruit.apo.ucla.edu/JPF1141…

This is a very good take. Clarity is no longer coupled with performance in math and theory research. Both are important, and operating at this pareto front could be our goal. We just need to adjust academic incentives to reward the whole curve a bit better.
We're clearly moving to a pareto front of "strenght of result" vs. "understandability" and as long as you are on that pareto front, you are all good.

We're clearly moving to a pareto front of "strenght of result" vs. "understandability" and as long as you are on that pareto front, you are all good.
Apparently, my coauthor pointed out that the OpenAI theorem drop included something we were working on (with AI help, of course). The OpenAI result has better bounds than we achieved, but it's also unreadable. Any advice on what I'm supposed to do now?

Given the sudden and amazing ubiquity of formal proofs in systems that implement the calculus of inductive constructions these days, when will Thierry Coquand, Christine Paulin-Mohring, and Gérard Huet get the Turing Award?

Cool research from @industriaalist. It's related to one of my favorite papers from @guyvdb arxiv.org/abs/2205.11502 which shows that transformers can represent many correct programs that backprop will never find. Below: a novel non-backprop based search mechanism.
Backprop has been the only credit assignment algorithm capable of training large neural nets. Introducing Dust, the first zeroth-order method to pretrain transformers to approach and sometimes even *exceed* backprop with large amounts of computation (bitter lesson!) - Dust is on the order of 1,000 to 10,000x more compute efficient than EGGROLL, the state-of-the-art ES method, for training transformers. - We chose pretraining transformers because it's arguably the hardest task possible for zeroth-order optimization. - Search in high dimensions is really misunderstood: zeroth-order usually gets better with overparameterization, not worse, i.e. at a fixed population size, larger models (up to 120x) reach a lower loss than smaller ones. - Backprop-like gradients emerge from a large population of activation perturbations, and the alignment holds up at every scale we tested, up to 1B tokens, which is critical for scaling. The core idea is a *virtual population*, which helps scale to large populations in transformers by perturbing activations in parallel. w/ @bishmdl76, @cs_serdar, @akshayvegesna

AI tools are very useful for mathematical research. But at present, producing good work requires a labor intensive step that I have been calling "deslopping" - making the thing readable. Please do not skip this step; even if the theorem is great an unreadable paper is worthless.

Making Google's Paper Assistant Tool available to ICLR submitters: blog.iclr.cc/2026/09/10/maki…

Terence Tao just posted this really insightful thread which seems to have been written in reaction to the Navier Stokes announcement. Please give it a read. mathstodon.xyz/@tao/11723732…

A Probabilistic Interpretation of KV Cache Eviction 🙇♂️🙇♂️🙇♂️🙇♂️🙇♂️
“A Probabilistic Interpretation of KV Cache Eviction” Most KV eviction methods use top-k heuristics to decide what to keep, but this can heavily distort attention after tokens are removed. This paper reframes KV eviction as expectation estimation, then uses probabilistic sampling and importance weighting to correct for missing KV entries during decoding. This gives a principled bias-variance tradeoff that stays competitive with existing methods while being more robust. alphaxiv.org/abs/2608.28293

What happens if you interpret attention scores as probabilities? We can then use sampling techniques in KV cache eviction. Check out our new paper!
“A Probabilistic Interpretation of KV Cache Eviction” Most KV eviction methods use top-k heuristics to decide what to keep, but this can heavily distort attention after tokens are removed. This paper reframes KV eviction as expectation estimation, then uses probabilistic sampling and importance weighting to correct for missing KV entries during decoding. This gives a principled bias-variance tradeoff that stays competitive with existing methods while being more robust. alphaxiv.org/abs/2608.28293

“A Probabilistic Interpretation of KV Cache Eviction” Most KV eviction methods use top-k heuristics to decide what to keep, but this can heavily distort attention after tokens are removed. This paper reframes KV eviction as expectation estimation, then uses probabilistic sampling and importance weighting to correct for missing KV entries during decoding. This gives a principled bias-variance tradeoff that stays competitive with existing methods while being more robust. alphaxiv.org/abs/2608.28293

web.cs.ucla.edu/~guyvdb/slid… for the slides