Longhui Yu (@scut_longhui) / X

Longhui Yu

445 posts

Longhui Yu

@scut_longhui

Post-train KIMI @Kimi_Moonshot | MS Peking University @PKU1898 Author of MetaMath, Easy2hard generalization, NuminaMath, Kimi k1.5, Kimi K2, K2 Thinking

Joined March 2018

Pinned
Longhui Yu
@scut_longhui
Jan 27
From Human Intelligence to Model Intelligence!🎯
Kimi.ai
@Kimi_Moonshot
Jan 27
🥝Meet Kimi K2.5, Open-Source Visual Agentic Intelligence. 🔹Global SOTA on Agentic Benchmarks: HLE full set (50.2%), BrowseComp (74.9%) 🔹Open-source SOTA on Vision and Coding: MMMU Pro (78.5%), VideoMMMU (86.6%), SWE-bench Verified (76.8%) 🔹Code with Taste: turn chats, images
497
Longhui Yu
@scut_longhui
Jul 13, 2025
Incredibly proud to have been part of this great project! I will never forget the night it was born—the anticipation and excitement we shared in the office! 😁😁😁
Kimi.ai
@Kimi_Moonshot
Jul 11, 2025
🚀 Hello, Kimi K2! Open-Source Agentic Model! 🔹 1T total / 32B active MoE model 🔹 SOTA on SWE Bench Verified, Tau2 & AceBench among open models 🔹Strong in coding and agentic tasks 🐤 Multimodal & thought-mode not supported for now With Kimi K2, advanced agentic intelligence
47K
Longhui Yu
@scut_longhui
Oct 22, 2023
🔥Mistral is really powerful! Introducing 🔥𝐌𝐞𝐭𝐚𝐌𝐚𝐭𝐡-𝐌𝐢𝐬𝐭𝐫𝐚𝐥-𝟕𝐁, trained on 𝐌𝐞𝐭𝐚𝐌𝐚𝐭𝐡𝐐𝐀 and achieved 𝟕𝟕.𝟕 on GSM8K (surpass all the 7B-13B models) and 𝟐𝟖.𝟐 on Math with COT only! Check at:
meta-math/MetaMath-Mistral-7B · Hugging Face
From huggingface.co
60K
Longhui Yu
@scut_longhui
Nov 20, 2023
🏅MetaMath-Llemma-7B, finetuned on MetaMathQA and based on Llemma-7B, achieves 69.2 on GSM8K and 30.0 on MATH. A really huge improvement on MATH. Thanks for your wonderful work! @zhangir_azerbay, @keirp1, @AlbertQJiang, @BlancheMinerva, @wellecks Check:
Longhui Yu
@scut_longhui
Oct 22, 2023
🔥Mistral is really powerful! Introducing 🔥𝐌𝐞𝐭𝐚𝐌𝐚𝐭𝐡-𝐌𝐢𝐬𝐭𝐫𝐚𝐥-𝟕𝐁, trained on 𝐌𝐞𝐭𝐚𝐌𝐚𝐭𝐡𝐐𝐀 and achieved 𝟕𝟕.𝟕 on GSM8K (surpass all the 7B-13B models) and 𝟐𝟖.𝟐 on Math with COT only! Check at: huggingface.co/meta-math/Meta…
meta-math/MetaMath-Llemma-7B · Hugging Face
From huggingface.co
9.5K
Longhui Yu
@scut_longhui
Jul 21, 2024
💡💡💡NuminaMath shows a dominant performance on AIMO, being the only solution to surpass the 25/50. The NuminaMath-CoT dataset introduces numerous new question-answer pairs, many of which were obtained by scanning PDFs (really valuable)!!! Huggingface: huggingface.co/AI-MO
1.6K
Longhui Yu
@scut_longhui
Mar 21, 2024
💡💡💡How can a model learn from easy tasks but perform beyond them? Is it possible to train a superhuman model? We investigate Easy-to-Hard Generalization on Math tasks and gain some insights. We hope these insights can drive research in scalable oversight and super alignment.
Zhiqing Sun
@EdwardSun0909
Mar 21, 2024
🌟Easy-to-Hard Generalization: Scalable Alignment Beyond Human Supervision 🌟 arxiv.org/abs/2403.09472 How can we keep improving AI systems when their capabilities surpass those of human supervisors? (1/n)
5.7K
Longhui Yu
@scut_longhui
Oct 22, 2023
Replying to @scut_longhui
To fine-tune Mistral-7B, I would suggest using a smaller learning rate (usually 1/5 to 1/10 of the lr for LlaMa-2-7B) and staying other training args unchanged. check our scripts at
github.com
GitHub - meta-math/MetaMath: MetaMath: Bootstrap Your Own Mathematical Questions for Large Language...
MetaMath: Bootstrap Your Own Mathematical Questions for Large Language Models - meta-math/MetaMath
1K
Longhui Yu
@scut_longhui
Jun 27, 2024
Really impressive！Just yesterday I was the top 1 on the leaderboard and I can't imagine how fast Lewis is improving really a good work！kaggle.com/competitions/a…
Lewis Tunstall
@_lewtun
Jun 26, 2024
Good data is all you need
1.4K
Longhui Yu
@scut_longhui
Jul 4, 2024
Numina is truly impressive. Before the deadline, we were initially leading Numina by 1 point. However, within just 1-2 days, Numina also reached 28 points and surpass us very soon! Moreover, I've heard this isn't even Numina's strongest version—there's an even more powerful one!!
Lewis Tunstall
@_lewtun
Jul 4, 2024
After 3 months of hard work, I'm heaps excited to share that our team won the first progress prize of the AI Math Olympiad 🥇! kaggle.com/competitions/a… This challenge involved fine-tuning open LLMs to solve 50 difficult math problems spanning geometry to number theory 🤓 Our
1.7K
Longhui Yu
@scut_longhui
Jun 27, 2024
Replying to @_lewtun
contrasts！iam the leader of tts.
703
Longhui Yu
@scut_longhui
Dec 29, 2023
Replying to @sybilhyz
Congrats!👏 The first paper I have seen that makes the Process reward Model work after Open AI's Lets verify step by step.
636
Longhui Yu
@scut_longhui
Jun 21, 2024
I am also currently visiting Westlake University, and it truly provides an exceptionally comfortable research environment.
Huan Wang
@huanwangx
Jun 14, 2024
Happy to share that I just started a new position as a tenure-track Assistant Professor @Westlake_Uni in June. Although switching from a phd student to a professor sounds a bit terrifying to me, I am particularly curious what kind of interesting and funny stories there will be!🤩
947
Longhui Yu
@scut_longhui
Nov 14, 2023
🏅We all know Orthogonality is very important in machine learning to keep something independent. In the era of Foundation Models, orthogonal finetuning can better maintain the pre-trained base knowledge and adapt to new tasks. Thanks to the awesome collaborators on BOFT!
Weiyang Liu
@Besteuler
Nov 14, 2023
📢Introducing BOFT: A New General Finetuning Method for the Adaptation of Foundation Models! Our latest research reveals that orthogonal finetuning is a versatile approach, effective across various tasks including vision, NLP, and text-to-image generation. 🧵1/5
1.2K
Longhui Yu
@scut_longhui
Nov 23, 2023
Rebuttal is nothing without reviewer's involvement
453