David Yin

Berkeley 25', First-year CS PhD student at Princeton University

[Richard Zhuang @RichardZ412]

How can we train small agentic models that are highly capable of terminal use and coding?

Announcing OpenThoughts-Agent + OpenThinkerAgent-32B, the strongest Qwen-3 based open-data agentic model: 44.8% avg across 7 agentic benchmarks! (1/n)

[Wenhao Chai @ COLM @wenhaocha1]

ByteDance has just released Seed 2.1, which achieved outstanding performance on two of our benchmarks: FrontierCS (https://t.co/1Oo1J0GttV) for frontier computer science problems, and WorldBench (https://t.co/BRNM4c5tsR) for multimodal knowledge.

Model Card: https://t.co/0H0wtCTw97

[Tony Chen @tonychenxyz]

Today, agents execute isolated tasks. Tomorrow, agents will steer complex decisions across long horizons.

Introducing CEO-Bench, a first step to measure "Steering Intelligence." In CEO-Bench, agents are asked to run a simulated startup for 500 days.

https://t.co/9gNWFALTKK

[Zhuang Liu @liuzhuang1234]

Can a small academic team build a strong text-to-image model using only public datasets?

Introducing i1: a simple, fully open recipe for strong text-to-image models

[Harish Krishnakumar @harishkrik]

Today’s vision benchmarks suggest VLMs are nearing saturation, but real-world visual understanding is far from solved.

Introducing WorldBench: 2,000 hand-written, human-verified VQA questions focused on visual diversity and designed to be challenging for frontier models. Gemini-3.1-Pro leads with just 64.0% accuracy. (1/10)

[Zhuang Liu @liuzhuang1234]

UEval: 1,000 questions, 10,417 rubric criteria, 8 tasks.

A challenging benchmark for the next generation of unified models.

Paper: https://t.co/Jk7xe4xPNQ
Dataset: https://t.co/aldylO9Uyl
Code: https://t.co/5p7BxEWZKS

Work led by Bo Li, and with @DavidYin0609, @wenhaocha1, @XingyuFu2.

[Zhuang Liu @liuzhuang1234]

We are ramping up - follow our lab GitHub account for more research and open-source releases in the new year!

https://t.co/rAxkxarOxe

[Zhuang Liu @liuzhuang1234]

Stronger Normalization-Free Transformers – new paper.

We introduce Derf (Dynamic erf), a simple point-wise layer that lets norm-free Transformers not only work, but actually outperform their normalized counterparts.

[David Yin @DavidYin0609]

H100s are just specific types of GPUs we have in one of the clusters at Princeton. It is actually way more than 0.8 GPUS/Students

[Francois Chaubard @FrancoisChauba1]

Last night, @agupta and I hosted a great dinner with 14 professors at #NeurIPS2025 from leading academic labs across the US, and many cited compute in academia as "abhorrent". Out of curiosity I just pulled these stats. This is insane. To do meaningful AI research today you need at least 1 GPU/student. Likely 8+ to be honest. The best university (Princeton) is at 0.8 GPUs/student. Stanford is at 0.14 GPUs/student. Marlowe (Stanford's "super cluster") has only 248 H100s for the whole CS Dept to use. Every frontier lab has >100k.

This needs to be fixed.

[Zhuang Liu @liuzhuang1234]

Excited to share our lab’s first open-source release: LLM-Distillation-JAX

supports practical knowledge distillation configurations (distillation strength, temperature, top-k/top-p), built on MaxText

designed for reproducible JAX/Flax training on both TPUs and GPUs

[Junlin Han @han_junlin]

Excited to share our new work: “Learning to See Before Seeing”! 🧠➡️👀 We investigate an interesting phenomeno: how do LLMs, trained only on text, learn about the visual world?
Project page: https://t.co/9mQt3qnckL

[David Yin @DavidYin0609]

Still remember earlier this year, I tried very hard to figure out why vLLM can give very different outputs than huggingface models even with greedy sampling (not sure if it is fixed now). Twisting batch size or number of gpus also makes the output different

[Thinking Machines @thinkymachines]

Today Thinking Machines Lab is launching our research blog, Connectionism. Our first blog post is “Defeating Nondeterminism in LLM Inference”

We believe that science is better when shared. Connectionism will cover topics as varied as our research is: from kernel numerics to prompt engineering. Here we share what we are working on and connect with the research community frequently and openly.

The name Connectionism is a throwback to an earlier era of AI; it was the name of the subfield in the 1980s that studied neural networks and their similarity to biological brains.

https://t.co/lrJioBmpbT

[Zhiqiu (Oscar) Xu @oscar_zhiqiu_xu]

How do we navigate a growing collection of post-trained LLMs?

In Delta Activations: A Representation for Finetuned LLMs, we propose a compact embedding that encodes the post-training signal.

Try the interactive model navigator 👉 https://t.co/EhxzGihVu5

[David Yin @DavidYin0609]

Check out our new paper “Generative Modeling of Weights: Generalization or Memorization?” — we find that current diffusion-based neural network weight generators often memorize training checkpoints rather than learning a truly generalizable weight distribution!

[Zhuang Liu @liuzhuang1234]

Can diffusion models appear to be learning, when they’re actually just memorizing the training data?

We show and investigate this phenomenon in the context of neural network weight generation, in our recent paper “Generative Modeling of Weights: Generalization or Memorization?"

[Lucas Beyer (bl16) @giffmana]

While we're at "recognizing evals", here is a legendary vision paper from CVPR 2011

It shows that it's quite easy to classify which dataset an image comes from (39% acc, vs random=8%).

The point being, every dataset having its distinct signature should be a default assumption.

[Roei Herzig @roeiherzig]

What an exciting #ICRA2025 week!✨

Catch our Oral & Poster this Wednesday in the Robotic Foundation Models Session — with @DavidYin0609, @zekaiw04 and @Dantong_Niu

Oral: Wed, 3:20 — 3:25pm, Room 314
Poster: Wed, 3:50 — 4:25pm, Room 314

Project: https://t.co/bGLv4oUJ7u

[David Yin @DavidYin0609]

Please come to our oral and poster presentations on Wednesday — I’ll be there along with @zekaiw04 and @Dantong_Niu to present our work!

Oral: Wednesday, 3:20 — 3:25pm, Room 314
Poster: Wednesday, 3:50 — 4:25pm, Room 314

Paper: https://t.co/VO2ZVwtFj4

[Roei Herzig @roeiherzig]

Have you wondered whether LLMs could directly execute robot tasks?🚨

Excited to share our work, *RoboPrompt*, a framework that enables off-the-shelf text-only LLMs to directly predict robot actions with ICL demonstrations without training! @berkeley_ai

https://t.co/bGLv4oUJ7u

[David Yin @DavidYin0609]

I had the same experience with ChatGPT. Once it misunderstands a question or gets something wrong, it will never go back to normal with more clarification.

Pro tip: open a new tab each time

[elvis @omarsar0]

LLMs Get Lost in Multi-turn Conversation

The cat is out of the bag.

Pay attention, devs.

This is one of the most common issues when building with LLMs today.

Glad there is now paper to share insights.

Here are my notes:

[Kenneth Li @ke_li_2021]

🧵1/
Everyone says toxic data = bad models.
But what if more toxic data could help us build less toxic models?
Our new paper explores this paradox. Here’s what we found 👇

[David Yin @DavidYin0609]

The post-training dataset is fully open!

[Oleksii Kuchaiev @ COLM2026 @kuchaev]

Llama-Nemotron-v1 technical report is now available on arxiv https://t.co/OwFdIZnYlH