Yi Xu (@_yixu) header
Yi Xu (@_yixu) profile picture

Yi Xu

@_yixu

Research @MistralAI | Previously @UCL_DARK

 London
yix8.github.io/
Joined June 2024
  • Tweets55
  • Following436
  • Followers628
Pinned Tweet
Yi Xu (@_yixu) profile picture
Yi Xu
@_yixuApr 10
🚀Scale a 1.6M-param transformer to one of today's strongest models, GPT-5.4, and it plans... 4 steps further. In our latest work, we measure how many steps of latent planning LLMs can discover and execute without chain-of-thought and find a persistent depth ceiling that resists scale. 🧵👇
Image from Yi Xu (@_yixu)
4
29
2
206
27,355
This was accepted to COLM 2026 🥳 A striking limitation of SotA LLMs in discovering a latent planning strategy in an almost trivial graph-search task Check out the paper ⤵️
Yi Xu (@_yixu) profile pictureYi Xu
@_yixuApr 10
🚀Scale a 1.6M-param transformer to one of today's strongest models, GPT-5.4, and it plans... 4 steps further. In our latest work, we measure how many steps of latent planning LLMs can discover and execute without chain-of-thought and find a persistent depth ceiling that resists scale. 🧵👇
Image from Yi Xu (@_yixu)
1
6
66
9,072
🚀 LLM from Trainee to Trainer. No manual environment redesign. No guessing where the model still struggles. 🤔   Let the model diagnose its own failures, and design the next training environment by itself. 🧠   We propose LLM-as-Environment Engineer, a new framework where LLMs not only learn in environments, but also learn to engineer the environments that teach them better. 🌍✨ The trainee becomes the trainer. And training becomes self-improving. 🚀
1
8
18
861
Happy to share that our paper Visual Planning has been selected as Oral for ICLR 2026 @iclr_conf 🥳🥳 #ICLR2026
Yi Xu (@_yixu) profile pictureYi Xu
@_yixu19 May 2025
🚀Let’s Think Only with Images. No language and No verbal thought.🤔 Let’s think through a sequence of images💭, like how humans picture steps in their minds🎨. We propose Visual Planning, a novel reasoning paradigm that enables models to reason purely through images.
Image from Yi Xu (@_yixu)
3
26
1
222
20,679
📘Full attention visualizations, error breakdowns, training dynamics, and ablations here! 📄Read the paper: arxiv.org/abs/2604.06427
The Depth Ceiling: On the Limits of Large Language Models in...

The Depth Ceiling: On the Limits of Large Language Models in...

The viability of chain-of-thought (CoT) monitoring hinges on models being unable to reason effectively in their latent representations. Yet little is known about the limits of such latent...

arxiv.org
1
2
21
741
Huge thanks to @philippjettkant and @LauraRuis for the incredible collaboration. It was an absolute joy working together!
10
571
Take a pause with textual CoT. 🧠✨ What if visual reasoning emerged from videos? Video models aren’t just generators. They can be generalizable, scalable reasoners. 🧵👇 Scaling! Scaling Test-Time Compute! More Frames = More Thinking Budget = Better Planning!
Image from Chengzu Li (@li_chengzu)
2
5
20
1,553
Self-Evolving AI Agents are about to blow up 🚀🔥 New Survey Alert 🚨 From static LLMs ➡️ lifelong, adaptive agents 🌱🤖 A Comprehensive Survey of Self-Evolving AI Agents... — your roadmap to the future of AI that learns & evolves after deployment! 💡 arxiv.org/pdf/2508.07407
Image from Zaiqiao Meng (@mengzaiqiao)
Image from Zaiqiao Meng (@mengzaiqiao)
2
14
4
38
4,660
LLMs can be programmed by backprop 🔎 In our new preprint, we show they can act as fuzzy program interpreters and databases. After being ‘programmed’ with next-token prediction, they can retrieve, evaluate, and even *compose* programs at test time, without seeing I/O examples.
Image from Laura Ruis (@LauraRuis)
4
57
7
312
34,796
Automating Multi-Agent Design: 🧩Multi-agent systems aren’t just about throwing more LLM agents together. 🛠️They require mastering the subtle art of prompting and agent orchestration. Introducing MASS🚀- Our new agent optimization framework for better prompts and topologies!
Image from Han Zhou (@hanzhou032)
14
160
10
714
81,847
🚀Let’s Think Only with Images. No language and No verbal thought.🤔 Let’s think through a sequence of images💭, like how humans picture steps in their minds🎨. We propose Visual Planning, a novel reasoning paradigm that enables models to reason purely through images.
Image from Yi Xu (@_yixu)
15
223
23
1,304
230,847
more replies
📘 Want to know more? 📄 Read the paper: arxiv.org/abs/2505.11409 ⭐️Star the repo: github.com/yix8/VisualPlanni…
Visual Planning: Let's Think Only with Images

Visual Planning: Let's Think Only with Images

Recent advancements in Large Language Models (LLMs) and their multimodal extensions (MLLMs) have substantially enhanced machine reasoning across diverse tasks. However, these models predominantly...

arxiv.org
2
5
34
3,305
Huge appreciation to @li_chengzu, @hanzhou032, and @wanxingchen_, @caiqizh, @annalkorhonen, @licwu for their incredible support and collaboration. It’s been truly rewarding to work together!
1
10
2,719
Load more