Jiayi Zhang (@didiforx) / X

Jiayi Zhang

353 posts

Jiayi Zhang

@didiforx

Ph.D. student @HKUSTGuangZhou, Researcher @MetaGPT_, Cofounder of OpenManus, previously at RUC, Lenovo Research AI Lab, Zhipu AI.

ShenZhen

didiforgithub.github.io

Joined July 2023

Jiayi Zhang
@didiforx
Mar 1, 2025
Reasoning models lack atomic thought ⚛️ Unlike humans using independent units, they store full histories🤔 Introducing Atom of Thoughts (AOT): lifts gpt-4o-mini to 80.6% F1 on HotpotQA, surpassing o3-mini and DeepSeek-R1 ! The best part? It's plugs in for ANY framework 🔌 1/5
396K
Jiayi Zhang
@didiforx
Mar 1, 2025
Replying to @didiforx
Want to try AOT? 📦 Code: github.com/qixucen/atom 📝 Paper: arxiv.org/abs/2502.12018 Huge thanks to main author @SteamedBun18755, co-authors @ZhaoyangYu22356, @AlexanderWu0, and the amazing @metagpt community for their support! ❤️ Follow us for more exciting updates! 5/5
GitHub - qixucen/atom: [NeurIPS 2025] Atom of Thoughts for Markov LLM Test-Time Scaling
From github.com
15K
Jiayi Zhang
@didiforx
Mar 6, 2025
No fortress, purely open ground. Manus 👋. We open-sourced its core feature in 2 hours after dinner. Check it out 👇: github.com/mannaandpoem/O… 1/4
00:00
00:00
75K
Jiayi Zhang
@didiforx
Mar 1, 2025
Replying to @didiforx and @SteamedBun18755
How does AOT work? ⚙️ For each reasoning step: 1. Decompose the question into DAG 2. Contract the subquestions into a NEW simpler question 3. Iterate until reaching an atomic question Just like a Markov process: each new question depends only on the previous state! 🎯 3/5
17K
Jiayi Zhang
@didiforx
Mar 1, 2025
Replying to @didiforx and @SteamedBun18755
Why do we need atomic thoughts? 🤔 ALL current reasoning approaches, both models (o3, R1...) and frameworks (CoT, ToT, GoT...) suffer from the same issue: keeping full reasoning histories. This leads to: Computationally expensive 💰 Prone to interference 🚫 2/5
20K
Jiayi Zhang
@didiforx
Mar 1, 2025
Replying to @didiforx
The power of a plug-in 🔌 AOT works with any approach: o3, R1, CoT, ToT, GoT, Self-Consistency, or Agentic Workflow. It simplifies inputs while preserving solution quality 💡 Try integrating AOT into your favorite approach! 4/5
15K
Jiayi Zhang
@didiforx
Mar 7, 2025
Replying to @dotey
宝玉老师，我们开源了一版Openmanus，能实现一部分功能。 github.com/mannaandpoem/O…
17K
Jiayi Zhang
@didiforx
Jun 19, 2025
It's actually a pity that we got no enough time to maintain OpenManus during the past 3 months. But the better news is that we will build a formal open-source community for OpenManus at the end of this month.
12K
Jiayi Zhang
@didiforx
Mar 2, 2025
Replying to @Comed_Ai_n and @SteamedBun18755
we have open sourced it in
GitHub - qixucen/atom: [NeurIPS 2025] Atom of Thoughts for Markov LLM Test-Time Scaling
From github.com
3K
Jiayi Zhang
@didiforx
Feb 17, 2025
No labels? ØvO help you! Excited to share our new paper: Self-Supervised Prompt Optimization (arxiv.org/abs/2502.06855) 🔥 Key features: ØvO: Output vs Output - no labels/human feedback needed! 99% cost reduction ($0.15) SOTA performance with just 3 examples 1/5
71K
Jiayi Zhang
@didiforx
Feb 14, 2025
Excited to share my first ICLR Oral Paper! Special thanks to @isaac_jinyu @ZhaoyangYu22356 @SteamedBun18755 @AlexanderWu0
MetaGPT
@MetaGPT_
Feb 14, 2025
🎉 AFLOW accepted for #ICLR2025 Oral! 🔧Easy to use for closed & open tasks! 📉Low inference costs with DeepSeek vs. larger models! ✨Promising research: Automatic Agentic Workflow/Systems! Paper: openreview.net/pdf?id=z5uVAKw… Code: github.com/geekan/MetaGPT…
16K
Jiayi Zhang
@didiforx
Mar 2, 2025
Replying to @LQGWarpSpeed and @SteamedBun18755
Maybe a repost will speed up the process hhh 😍
6.6K
Jiayi Zhang
@didiforx
Mar 8, 2025
Important Announcement ⚠ We have noticed that certain accounts are impersonating our team and claiming to launch a token called "$OpenManus". We hereby declare: 1. OpenManus is a legitimate project developed by the MetaGPT team 2. We have NEVER issued any cryptocurrencies
GitHub - mannaandpoem/OpenManus
From github.com
4.5K
Jiayi Zhang
@didiforx
Mar 18, 2025
Text-to-SQL woes? Reasoning models stumble in zero-shot tasks 😓 Enter Alpha-SQL — our breakthrough boosts 7B LLMs by 15-20%, topping GPT-4o SOTA and even reasoning models on BIRD! 🎉 Test Time Scaling still shines. How we nailed it 👇: 1/5
2.6K