Can LLMs reason beyond context limits? 🤔 Introducing Knowledge Flow, a training-free method that helped gpt-oss-120b & Qwen3-235B achieve 100% on the AIME-25, no tools. How? like human deliberation, for LLMs. 📝 Blog: yufanzhuang.notion.site/knowledge-flow 💻 Code: github.com/EvanZhuang/kno…

Check our Knowledge Flow blog: We develop a new axis of test-time scaling by doing iterative refinement on a "knowledge" list for reasoning tasks! Notably, we find that updating what is wrong is more effective than recording what is right. Great job led by Yufan Zhuang!

got confused by something basic and went down a rabbit hole, so I just wrote a blogpost about it. "Is Density vs. Feature Coverage That Different?" nanjiang.cs.illinois.edu/2025/10/24/cov…

Our latest post explores on-policy distillation, a training approach that unites the error-correcting relevance of RL with the reward density of SFT. When training it for math reasoning and as an internal chat assistant, we find that on-policy distillation can outperform other

Today we’re announcing research and teaching grants for Tinker: credits for scholars and students to fine-tune and experiment with open-weight LLMs. Read more and apply at: thinkingmachines.ai/blog/tinker-re…

this would be a very good starting point for learning / prototyping many times people interested in learning ML/DL/LLM are intimidated by the sys/compute complexity

🚀Excited to share our new work! 💊Problem: The BF16 precision causes a large training-inference mismatch, leading to unstable RL training. 💡Solution: Just switch to FP16. 🎯That's it. 📰Paper: arxiv.org/pdf/2510.26788 ⭐️Code: github.com/sail-sg/Precis…

Daniel Han, glad you liked the post! You're spot on to suspect lower-level implementation issues. That's exactly what we found in the original blog. The disable_cascade_attn finding (Sec 4.2.4) was the symptom, but the root cause was that silent FlashAttention-2 kernel bug

Yingru Li Daniel Han Hi Yingru Li , I tried this disable_cascade_attn many times, including the latest vllm version. But unfortunately it made no difference in our experiments. So I guess it really depends on the setting.

Life update: I will join NYU Courant Math (CAOS) and Center of Data Science as an assistant professor in Fall 2026. If you are interested in doing a PhD with me please let me know! zongyi-li.github.io

It’s frustrating how labs like Kimi.ai, Qwen, DeepSeek, Ai2, Hugging Face... share their research, pipelines, and lessons openly, only for closed-source labs to quietly use that knowledge to build better models without ever giving back.

Science is better when shared!

We do! It's indeed better & very interesting! Check x.com/yufan_zhuang/s…

"Dense Backpropagation Improves Pretraining for sMoEs" is accepted at NeurIPS Conference! We show that we can proxy inactive experts with a cheap estimator, and that doing this in pretraining improves performance without requiring HPO or compute overhead.

RL is bounded by finite data😣? Introducing RLVE: RL with Adaptive Verifiable Environments We scale RL with data procedurally generated from 400 envs dynamically adapting to the trained model 💡find supervision signals right at the LM capability frontier + scale them 🔗in🧵

🚀 No More Train–Inference Mismatch! We demonstrate bitwise consistent on-policy RL with TorchTitan (training) + vLLM (inference) — the first open-source run where training and inference numerics match exactly. It only takes 3 steps: 1️⃣ Make vLLM batch-invariant (same seq →

I’ve been thinking a lot about what the net benefit of the AI platform wave is. The real question is how to empower every company out there to get more out of this platform shift and build their own AI native capabilities and enterprise value (vs inadvertently just transfer their

Congratulations!

Congratulations!