Pinned Tweet
The universe runs toward disorder. Yet out of it came life, then minds, and now AI. Why?
The twist: order speeds up disorder. Life burns energy faster than rock. Humans tapped coal and uranium. AI may be next.
Opinions are mine. Pushbacks are welcome! x.com/i/article/2111…
The twist: order speeds up disorder. Life burns energy faster than rock. Humans tapped coal and uranium. AI may be next.
Opinions are mine. Pushbacks are welcome! x.com/i/article/2111…
7
2
26
4,175
Code agents keep surprising us with what they can do in embodied tasks. But the world doesn’t pause while they think.
Whether it’s a code agent or a VLA, a policy needs to account for time and dynamics, including how the world changes while it plans.
This is what we study in LAGEN: latency awareness, an important but often overlooked capability for embodied agents.
Huge thanks to Zihan for leading this effort and to the whole team for the amazing work!
Whether it’s a code agent or a VLA, a policy needs to account for time and dynamics, including how the world changes while it plans.
This is what we study in LAGEN: latency awareness, an important but often overlooked capability for embodied agents.
Huge thanks to Zihan for leading this effort and to the whole team for the amazing work!
Introducing LAGEN: Latency-Aware Agents.
A 1-frame delay is invisible to humans, but that delay can make an agent lose 95% performance on real-time tasks.
AI models like Astra and VLAs can already get to real-time loops. But to keep them alive, we need Latency Awareness.🧵
1
3
15
919
Qineng Wang retweeted
Introducing LAGEN: Latency-Aware Agents.
A 1-frame delay is invisible to humans, but that delay can make an agent lose 95% performance on real-time tasks.
AI models like Astra and VLAs can already get to real-time loops. But to keep them alive, we need Latency Awareness.🧵
A 1-frame delay is invisible to humans, but that delay can make an agent lose 95% performance on real-time tasks.
AI models like Astra and VLAs can already get to real-time loops. But to keep them alive, we need Latency Awareness.🧵
7
18
102
24,138
Qineng Wang retweeted
Huge congratulations on the work from our lab! Really exciting to see this.
I've been thinking about why embodied AI remains challenging, and where the next breakthroughs will come from. AI has repeatedly taught us: many seemingly architectural problems are indeed data problems. And this makes embodied intelligence so interesting.
Data from physical world is hard to get. We cannot simply scrape billions robots interaction trajectories like web pages. But if we manage to have that embodied data, even models like astra originally built for language tasks can potentially solve embodied tasks.
So I increasingly believe the next frontier of AI will be defined as much by where data is missing as by where model capabilities are missing.
And this should not stop at robotics. Much of the world's intelligence exists as tacit knowledge: the hands of experienced technicians, the judgment of nurses, the intuition of farmers, and countless small decisions people make every day to keep the physical world running.
Frontier models becomes more capable at embodied tasks, which we should celebrate. But a this doesn't mean that embodied intelligence is nearly solved, or that the remaining challenges will automatically disappear with the next model release.
This raises a broader question: how much of the world's intelligence have we actually captured in a form AI can learn from?
It's easy to forget how many people have never paid for AI, especially when we're surrounded by people who hit their coding agent usage limits every 5 hours.
Some of the most interesting challenges lie in environments where collecting a single useful example is harder than training on a million existing ones.
That's what makes embodied intelligence so interesting, and why I'm excited to see Embodied First Exam helping us understand what's still missing.
Congratulations again to the team. Excited to see where this goes!
I've been thinking about why embodied AI remains challenging, and where the next breakthroughs will come from. AI has repeatedly taught us: many seemingly architectural problems are indeed data problems. And this makes embodied intelligence so interesting.
Data from physical world is hard to get. We cannot simply scrape billions robots interaction trajectories like web pages. But if we manage to have that embodied data, even models like astra originally built for language tasks can potentially solve embodied tasks.
So I increasingly believe the next frontier of AI will be defined as much by where data is missing as by where model capabilities are missing.
And this should not stop at robotics. Much of the world's intelligence exists as tacit knowledge: the hands of experienced technicians, the judgment of nurses, the intuition of farmers, and countless small decisions people make every day to keep the physical world running.
Frontier models becomes more capable at embodied tasks, which we should celebrate. But a this doesn't mean that embodied intelligence is nearly solved, or that the remaining challenges will automatically disappear with the next model release.
This raises a broader question: how much of the world's intelligence have we actually captured in a form AI can learn from?
It's easy to forget how many people have never paid for AI, especially when we're surrounded by people who hit their coding agent usage limits every 5 hours.
Some of the most interesting challenges lie in environments where collecting a single useful example is harder than training on a million existing ones.
That's what makes embodied intelligence so interesting, and why I'm excited to see Embodied First Exam helping us understand what's still missing.
Congratulations again to the team. Excited to see where this goes!
Introducing Embodied First Exam:
One line of code to run coding agents in 1,090 embodied tasks across 31 simulation suites and 44 robots.
Can your agent go far beyond pick-and-place:
Puzzles with spatial constraints;
Olympic sports with multi-robot collab;
Real-time eval, where Dynamic, agile, whole-body control;
AutoResearch with sim2sim
Try your own agent:
harbor run -d embodied-first-exam -a <your_own_harness> -m <your_own_model>
0
6
29
2,598
Qineng Wang retweeted
🚀 We’re hiring at MAGIC Lab, NUS!
My lab is recruiting 1 postdoctoral researcher, 2 PhD students, and 3 research assistants to join our team and work on ambitious robotics projects!
Singapore’s robotics ecosystem is growing rapidly, with companies and organizations such as NVIDIA Singapore, the Singapore Embodied AI Centre, Sharpa, and many others contributing to an increasingly vibrant research community.
There has never been a more exciting time to work on robotics in Singapore!
Apply here: https://
magiclab-nus.github.io
If you’re interested in joining us, please reach out and feel free to share this with anyone who might be a good fit!
My lab is recruiting 1 postdoctoral researcher, 2 PhD students, and 3 research assistants to join our team and work on ambitious robotics projects!
Singapore’s robotics ecosystem is growing rapidly, with companies and organizations such as NVIDIA Singapore, the Singapore Embodied AI Centre, Sharpa, and many others contributing to an increasingly vibrant research community.
There has never been a more exciting time to work on robotics in Singapore!
Apply here: https://
magiclab-nus.github.io
If you’re interested in joining us, please reach out and feel free to share this with anyone who might be a good fit!
5
29
304
12,974
Coding agents have demonstrated remarkable capabilities. But where does their frontier lie in the physical world?
Excited to have helped build Embodied First Exam, where we probe the boundaries of general-purpose agents across a wide range of physical challenges: intricate spatial constraints, precision painting, dynamics-heavy RoboOlympics sports, dexterous manipulation, sim-to-real gaps, and much more.
It's fascinating to see what these agents can already do, and where they still struggle. Lots more to explore!
Excited to have helped build Embodied First Exam, where we probe the boundaries of general-purpose agents across a wide range of physical challenges: intricate spatial constraints, precision painting, dynamics-heavy RoboOlympics sports, dexterous manipulation, sim-to-real gaps, and much more.
It's fascinating to see what these agents can already do, and where they still struggle. Lots more to explore!
Introducing Embodied First Exam:
One line of code to run coding agents in 1,090 embodied tasks across 31 simulation suites and 44 robots.
Can your agent go far beyond pick-and-place:
Puzzles with spatial constraints;
Olympic sports with multi-robot collab;
Real-time eval, where Dynamic, agile, whole-body control;
AutoResearch with sim2sim
Try your own agent:
harbor run -d embodied-first-exam -a <your_own_harness> -m <your_own_model>
1
1
37
2,091
I've thought about this for a long time and finally wrote it up. Opinions are mine. Claude Opus 5.5, GPT 6.1 Solar, and Gemini 4 helped with examples, fact-checks, and translation.
0
0
1
85
宇宙总在走向无序，偏偏诞生了生命，又演化出了人类。而现在，人类正在造出可能更智能的 AI。为什么？
答案有点反直觉：越有序的东西，越能加快宇宙走向无序。生命比石头更会消耗能量，人类靠知识用上了煤和铀。接下来，也许轮到 AI。
仅代表个人观点，欢迎讨论！ x.com/i/article/2111…
答案有点反直觉：越有序的东西，越能加快宇宙走向无序。生命比石头更会消耗能量，人类靠知识用上了煤和铀。接下来，也许轮到 AI。
仅代表个人观点，欢迎讨论！ x.com/i/article/2111…
6
4
12
8,638
Qineng Wang retweeted
🚀 We’re hiring at MAGIC Lab, NUS!
My lab is recruiting postdoctoral researchers, PhD students, research assistants, and research engineers to join our growing team and work on ambitious robotics projects.
Singapore’s robotics ecosystem is growing, with companies such as NVIDIA Singapore, Sharpa, Flexiv, and AMI, alongside an expanding community of researchers. It’s an exciting time to be working in robotics here!
Apply here: https://
magiclab-nus.github.io
If you’re interested in joining us, please reach out and feel free to share this with anyone who might be a good fit!
My lab is recruiting postdoctoral researchers, PhD students, research assistants, and research engineers to join our growing team and work on ambitious robotics projects.
Singapore’s robotics ecosystem is growing, with companies such as NVIDIA Singapore, Sharpa, Flexiv, and AMI, alongside an expanding community of researchers. It’s an exciting time to be working in robotics here!
Apply here: https://
magiclab-nus.github.io
If you’re interested in joining us, please reach out and feel free to share this with anyone who might be a good fit!
10
34
408
20,578
I’ve always found the gap between video generative models and VLMs interesting. One learns how the world moves, while the other is much better at talking and reasoning about it. In our new work, we ask whether video generative models can help VLMs learn to reason about how the world changes👇
🚨 What makes good world-model training for MLLMs that actually transfers across tasks?
🧶 Excited to share our new work, WOVEN!
🧩 Spatial, physical, temporal, and embodied reasoning all require understanding how the world changes.
🔁 We view world modeling through state ), and study visual transition reasoning: inferring what is unobserved in (s, a, s′).
🤔 Can visual transition reasoning be learned as a shared primitive that benefits multiple downstream tasks?
🤔 If so, how should we train it for better downstream gains?
Major takeaways:
➡️ Visual transition reasoning is a shared, teachable primitive that transfers widely downstream: ~2K-example subsets collectively improve 22 of 26 downstream benchmarks by up to 27.3 points.
🧪 A data-centric training recipe: to teach this shared primitive for downstream tasks, select supervision by the reasoning operation it teaches, rather than scene/action/domain similarity.
Thread🧵👇
1
0
9
774
Qineng Wang retweeted
🚨 What makes good world-model training for MLLMs that actually transfers across tasks?
🧶 Excited to share our new work, WOVEN!
🧩 Spatial, physical, temporal, and embodied reasoning all require understanding how the world changes.
🔁 We view world modeling through state ), and study visual transition reasoning: inferring what is unobserved in (s, a, s′).
🤔 Can visual transition reasoning be learned as a shared primitive that benefits multiple downstream tasks?
🤔 If so, how should we train it for better downstream gains?
Major takeaways:
➡️ Visual transition reasoning is a shared, teachable primitive that transfers widely downstream: ~2K-example subsets collectively improve 22 of 26 downstream benchmarks by up to 27.3 points.
🧪 A data-centric training recipe: to teach this shared primitive for downstream tasks, select supervision by the reasoning operation it teaches, rather than scene/action/domain similarity.
Thread🧵👇
🧶 Excited to share our new work, WOVEN!
🧩 Spatial, physical, temporal, and embodied reasoning all require understanding how the world changes.
🔁 We view world modeling through state ), and study visual transition reasoning: inferring what is unobserved in (s, a, s′).
🤔 Can visual transition reasoning be learned as a shared primitive that benefits multiple downstream tasks?
🤔 If so, how should we train it for better downstream gains?
Major takeaways:
➡️ Visual transition reasoning is a shared, teachable primitive that transfers widely downstream: ~2K-example subsets collectively improve 22 of 26 downstream benchmarks by up to 27.3 points.
🧪 A data-centric training recipe: to teach this shared primitive for downstream tasks, select supervision by the reasoning operation it teaches, rather than scene/action/domain similarity.
Thread🧵👇
2
9
21
3,120
Qineng Wang retweeted
Truly appreciate it!
@ManlingLi_ will hold the Kuruganti Junior Faculty Fellowship for a one-year term. Established through the generosity of Adi Kuruganti (’99, MBA ’06), the fellowship supports early-career faculty working in computational research related to AI and ML.
spr.ly/6013BGvdB5
14
8
149
10,984
Qineng Wang retweeted
@ManlingLi_ will hold the Kuruganti Junior Faculty Fellowship for a one-year term. Established through the generosity of Adi Kuruganti (’99, MBA ’06), the fellowship supports early-career faculty working in computational research related to AI and ML.
spr.ly/6013BGvdB5
spr.ly/6013BGvdB5
0
4
20
9,403
Qineng Wang retweeted
Astra already shows exciting potential for real2sim2real for solving robotics tasks. Our DexAgent harness takes it further: no task-specific robot demos, 3.9× the success rate of zero-shot Astra across 11 real-world tasks (63.6% vs. 16.4%)! Check it out!
Learning bimanual dexterous manipulation from a single human video? Agents can be a solution!
Introducing DexAgent, an agentic Human2Sim2Robot framework for dexterous manipulation with a self-evolving tool library.
We evaluated DexAgent on 11 real-world, long-horizon dexterous ning rigid, articulated, and deformable objects.
🤖 63.6% policy rollout success rate, compared with 18.2% for the strongest baseline
⚡ 2.1 hours average inference time with the tool library, compared with 3.3 hours for the baseline
🧰 A self-evolving library with 103 skills + 188 verifiers, designed to keep growing as DexAgent encounters new tasks and objects
Learn more and contribute to the growing tool library:
Project Website: dexagent1.github.io
Arxiv: arxiv.org/abs/2609.35318
1
4
49
5,923
GPT-6 Astra can solve some seriously tricky spatial constraint tasks: from unlocking interlocked parts to threading a rope through three rings.
Try it yourself now 👇
qinengwang-aiden.github.io/demos/constrai…
Try it yourself now 👇
qinengwang-aiden.github.io/demos/constrai…
64
216
2,514
649,244
Qineng Wang retweeted
For the first time, scientists have mapped the complete brain and central nervous system of an adult male fruit fly — a key model organism in science. 🪰
Working alongside HHMI Janelia Research Campus and the scientific community, @GoogleResearch scientists and researchers used of 2D images into 3D neural shapes, reconstructing a record-breaking 166,000+ neurons. This foundational map of the adult male fruit fly brain can help accelerate our understanding of the brain, and is a major milestone in neuroscience.
Working alongside HHMI Janelia Research Campus and the scientific community, @GoogleResearch scientists and researchers used of 2D images into 3D neural shapes, reconstructing a record-breaking 166,000+ neurons. This foundational map of the adult male fruit fly brain can help accelerate our understanding of the brain, and is a major milestone in neuroscience.
452
3,010
30,294
16,829,787
Qineng Wang retweeted
I trained a controller for Microduck using a 192-neuron Male CNS connectome subset. Watch its modeled neural activity drive exploratory head and body movements. Fun stuff.
25
73
608
33,880
Qineng Wang retweeted
Today is the day! 🎯 Failure Modes in Agentic AI (FAGEN) Workshop @ ICML 2026.
Nine speakers, 218 submissions, one shared question: what does it mean for an agent to fail, and how do we fix it?
Find it. Reproduce it. Trace it. Fix it.
📍 COEX Ballroom 104–105 · all day. See you there!
Nine speakers, 218 submissions, one shared question: what does it mean for an agent to fail, and how do we fix it?
Find it. Reproduce it. Trace it. Fix it.
📍 COEX Ballroom 104–105 · all day. See you there!
4
9
49
11,517