I want a cafe, but for protein shakes

MathArena goes visual: We evaluated models such as GPT-5 on Math Kangaroo 2025, a recent contest for ages 6-19 where most tasks require visual reasoning. Models struggle the most with tasks for younger kids. For example, they get this task for 1st graders only 3% of the time 🧵

It's a shame, because I really like the approach of this paper, but why did they include a prompt injection in the arXiv version? public shame :(

cool stuff! personally very frustrated by hidden bugs and debugging unwanted changes in diffs

Niklas Sheth Ron Arel Intology their 'superhuman' ai cleverly assigned all the work to non-default streams, which means the correctness test (which waits on all streams) passes, while the profiling timer (which only waits on the default stream) is tricked into reporting a huge speedup

I will attend #neurips2025 chat me up about - llms for code - constrained decoding - diffusion LLMs

two weeks into #iclr rebuttal, so far 1 desk reject, 1 withdrawal and 1 reviewed paper that actually submitted a rebuttal :( to be clear, the remainder are *not* clear accepts.

there are many good reasons to reject a paper anonymously there are no good reasons to change your score after being deanonymized

I'll be in SF for a week from Dec. 8, and would love to learn about any and all problems you are facing when using LLMs (in particular for code). DM me if you'd like to grab an ice cream and chat 🍨