Zizhao Chen

Today’s AI models can’t even tie their own shoes. New research—led by @ch272h—tests AI models in a 3D environment, finding they perform well at untangling basic knots but cannot tie knots from simple loops or convert one knot to another. @Cornell_Bowers news.cornell.edu/stories/2025/1…

Merriam-Webster’s human editors have chosen ‘slop’ as the 2025 Word of the Year.

@yoavartzi I'm presenting on Friday. Details below: Fri, Dec 5, 2025 11:00 AM – 2:00 PM PST Exhibit Hall C,D,E #4505 Pic: knots inside USS midway museum near SD convention center

✨ Why it matters KnotGym gives us a lightweight yet expressive testbed for multi-modal long-horizon reasoning and planning. 🔗 Website: lil-lab.github.io/knotgym 📄 Paper: arxiv.org/abs/2505.18028 Joint work with @yoavartzi

🧩Natural language isn’t all you need. We’re great at evaluating text-based reasoning (MATH, AIME…) but what about long-horizon visual reasoning? Enter 𝗞𝗻𝗼𝘁𝗚𝘆𝗺: a minimalistic testbed for evaluating agents on spatial reasoning along a difficulty ladder

@denisparra neurips.cc/virtual/2025/l… Exhibit Hall C,D,E #4505 Fri 5 Dec 11 a.m. PST — 2 p.m. PST

@ch272h interesting! which day and session, and which poster number do you have ?

Hi all, I will be at #NeurIPS2025 to present my work on stress-testing looooooong visual reasoning with KnotGym🥨 Let's talk, whether or not your VLM that can see 14 million possible futures like Doctor Strange

🧵 New paper: "Simple Context Compression" - we show that mean-pooling beats the widely-used compression-tokens method for compressing contexts in LLMs, while being simpler and more efficient! with @yoavartzi (1/7)

Pushed a big update to LM-class (v2025.2) -- this second version makes a much more mature resource Many refinements of lecture slides + significant improvements to the assignments Many thanks to @ch272h @HuaYilun and @shankarpad8 for their work on the assignments

🚨Modeling Abstention via Selective Help-seeking LLMs learn to use search tools to answer questions they would otherwise hallucinate on. But can this also teach them what they know vs not? @momergul_ introduces MASH that trains LLMs for search and gets abstentions for free! 💡Key idea: Reward accuracy but penalize searches during training. Under the right optimization pressure, LLMs learn to invoke search when their parametric knowledge is lacking. At inference, we simply remove this search access and treat any search invocation as a proxy for abstention!

ToddlerBot 2.0 is released🥳! Now Toddy can also do cartwheels🤸! We have added so many features since our first release in February; see github.com/hshi74/toddler… for more details. Threads🧵(1/n)

@yoavartzi @xhluca @giffmana And there are cool things like dependent types

@xhluca @giffmana True. But because it was added on top of a thriving language, someone had to decide either to alienate the entire world or make it optional and mild. They correctly choose the latter route. Java was type-safe-first, and that makes for a very different beast

I love codex and claude taking care of all the boilerplate part of coding that wastes time and is booooooring. Come to think of it, maybe Java would in theory be the perfect language for LLM-coding? Extremely verbose boilerplate - very annoying for human, but good for LLM?

@yoavartzi @xhluca @giffmana did you know moonbit? #seamless-integration-of-the-python-ecosystem" target="_blank" rel="nofollow noopener">moonbitlang.com/blog/moonbit-x…

@xhluca @giffmana I do have this internal bet that we will see a prog. lang. that is built for LLMs-first coming up at some point. It will be interesting. But then there's the chicken-and-egg problem of data

@YouJiacheng @Pushpendre89 @yoavartzi … to train a 350M model. it’s all good btw

@Pushpendre89 @yoavartzi cuz Yoav just occupied the node lol

Me: the new GPU node is online My students: 💃🕺💃🕺💃 Me: torchrun --standalone --nproc_per_node=8 train.py My students: 🤬🤬🤬🤬🤬

@_xjdr where were they stuck?

all 3 gave up in less than 12 hours with nothing resembling a functional forward pass. feels like there is still a giant gap here

@yoavartzi Procrastination gets you anything but :(

@ch272h digs the best gems! recnet.io/rec/c863a1ca-9…

Why scaling data annotators at high costs, when you can scale users for free?

Time to democratize humanoid robots! Introducing ToddlerBot, a low-cost ($6K), open-source humanoid for robotics and AI research. Watch two ToddlerBots seamlessly chain their loco-manipulation skills to collaborate in tidying up after a toy session. toddlerbot.github.io