Mellen Y. Pu

💎Introduce PiEvo, the "Brain" of Agentic AI for Scientific Discovery.

[Mellen Y. Pu @Melletenmoon]

Excited to present #PiEvo !

📍 Poster session @ Wed, Jul 8, 2026 • 5:00 PM – 6:45 PM, KST HALL A #902

In the meantime, feel free to reach out if you have any questions — happy to discuss!

See you at #ICML!

[Mellen Y. Pu @Melletenmoon]

@hanminwang3 @my_cat_can_code Good question..

[Mellen Y. Pu @Melletenmoon]

@pranamanam really nice work, AFE! Congratulations!

[Mellen Y. Pu @Melletenmoon]

@vista8 能用其他llm endpoint吗

[Mellen Y. Pu @Melletenmoon]

@NFT_Chen 这么快，但这会在准确度上存在问���吗？

[Mellen Y. Pu @Melletenmoon]

I completely fail to understand the motivation behind this... If it were possible to track LLM cheating and hallucinations within the cache, that would indeed be cool...

[AYi @AYi_AInotes]

清华大学和无问芯穹等团队刚刚开源了一篇注定载入史册的重磅论文（��收录于 ICLR 2026，代码已开源）：
他们彻底打破了所有多智能体（Multi-Agent）系统的底层范式——
大型语言模型之间，从今天起可以彻底不通过人类文字进行直接交流了！

这项被称为 C2C（Cache-to-Cache，缓存到缓存）的技术，
开篇就指出了全网 Agent 协同里最愚蠢、也最昂贵的一个死穴：
为什么两个跑在同一台服务器里的超级 AI，交流时非要像人类一样，
把脑子里极其丰富的高维思考，硬生生压缩成一个个英文或中文单词打给对方看？

现有多 Agent 协同的痛苦，做过工程的人都懂：
模型 A 思考完，必须经历极其���长、吃显存带宽的逐字解码（Decode），慢吞吞吐出一大堆文字；
模型 B 拿到这堆文字，再重新过一遍分词和前缀计算。
这一来一回，不仅丢失了海量微观的语义细节，更造成了毁灭级的延迟和 Token 账单！

清华团队这次做了一件极其惊艳的体系化反叛——C2C 脑电波直连：

1️⃣ 彻底消灭中间文本，直接做“思维投影”：
模型 A 算完后，根本不生成任何文字，
系统直接用一个轻量级神经网络（Neural Fuser），把模型 A 的注意力记忆（KV-Cache）通过高维空间旋转和对齐，直接嫁接、融进模型 B 的 KV-Cache 脑内！
这相当于人类交流跳过了声带和耳朵，直接把我的记忆切片瞬间印进你的��脑；
2️⃣ 可学习的智能门控（Gate 机制）：
不同模型的网络层数和结构各不相同，直接硬灌肯定会神经错乱。
论文设计了一套精妙的可学习门控：
大模型会自己动态感知，精确挑选哪些关键层吸收外部缓存的增益最高，哪一层该保持独立思考，毫秒级自适应配平；
3️⃣ 极其残暴的工业基准成绩：
• 彻底消灭中间所有的文本解码等待，推理速度直接飙升 2.0 到 2.5 倍；
• 综合问答准确率比单模型最高暴涨 14.2%；
• 最狠的是：它的表现比传统用文字互相聊天的 Agent 团队，还要高出整整 5%！
因为在直接的语义融合下，原本会被文字丢弃的逻辑细微特征，全部被完整保真保留了。

这篇论文最刺痛人的一层哲学启示在于：
人类一直自豪地以为，自然语言是这个星球上传递智慧最完美的介质；
但在机器与机器的交互世界里，人类语言不过是一道充满歧义、极其拥挤的低速单行道。

当两个模型不再需要通过文字向对方妥协，
它们正在绕过人类的喉咙，
在冰冷的显存带宽之间，编织出一座全人类根本听不见、却快得不可思议的超高速无声神经网络。

[Mellen Y. Pu @Melletenmoon]

@AKurian001 amazing! are they OOD samples?

[Mellen Y. Pu @Melletenmoon]

For recent news about OpenAI and NS-equation,

1. estimated cost ranges from $10 million to $15 million.
2. still a long road ahead between making a claim.

We may have obtained the correct answer, yet lost the process of understanding it.

[Mellen Y. Pu @Melletenmoon]

@The_Alex Crazy……

[Mellen Y. Pu @Melletenmoon]

@sirbayes Really nice. Could this be used for discovering some laws in science? I think the law in science must be explicit and, specifically well-validated... Thanks!

[Mellen Y. Pu @Melletenmoon]

Really worth reading if you want to touch graph theory, thanks for sharing!

[Kirk Borne @KirkDBorne]

Introduction to Graph Theory: https://t.co/doAI3c6XgT (454-page PDF) — edges, vertices, trees, networks, paths, and graphs (and theorems!)
—————
#Mathematics #NetworkScience

[Mellen Y. Pu @Melletenmoon]

@maximilien912 impossible. I don’t believe Gemini is good enough for daily use

[Mellen Y. Pu @Melletenmoon]

@TaNGSoFT 值不值钱, 也许需要重新定义

[Mellen Y. Pu @Melletenmoon]

@sirbayes https://t.co/kzLoliWe6l

[Mellen Y. Pu @Melletenmoon]

@sirbayes looks insightful

[Mellen Y. Pu @Melletenmoon]

@chenru_duan sounds cool.

[Mellen Y. Pu @Melletenmoon]

@YiMaTweets Agree. The process of people creating knowledge is also based on external feedback. The so-called end-to-end approach simply closes off knowledge at the beginning – which is obviously a limitation.

[Mellen Y. Pu @Melletenmoon]

AI Scientist! 🧑‍🔬

[Jeff Dean @JeffDean]

Announcing Discovery Loop!

I am very excited to announce that, along with my longtime friends and collaborators @Sanjay_Ghemawat, @OriolVinyalsML and @quocleix, we are founding Discovery Loop (@DiscoLoopAI), a Public Benefit Corporation whose mission is to automate machine learning, science, and engineering to accelerate discoveries and progress. The four of us have worked together for 14 to 30 years, and have helped build some of the world’s most used products, infrastructure and AI models, and we’re excited to turn our attention to this ambitious endeavor.

♾

Learn more at: https://t.co/Rv3LMdLluK

[Mellen Y. Pu @Melletenmoon]

@zhongerxin @annaxtime 支持提前登��账号并缓存等功能吗？

[Mellen Y. Pu @Melletenmoon]

People often place excessively high expectations on new technologies. For example, some say that an LLM is about studying physics, but those who say this don't even understand what physics is. Mere similarity is meaningless.

[Alexander Yue @Alezander9]

Another week of Opus 5 making this game and now it has a full snow simulation

[Mellen Y. Pu @Melletenmoon]

Scientific discovery is like finding a pathway towards the other side of a forest. One should hypothesize the underlying “map” of science — how does it work internally. Verify it, then use it.