Reasoning agent / RL / efficiency research @NorthwesternU & @nvidia. Ex @Microsoft @yutori_ai @deepseek_ai @uiuc_nlp @RUC1937.
- Tweets 1,342
- Following 700
- Followers 22,951
- Likes 4,693
Pinned Tweet
🧵 Claude-Opus-4.8 takes you too much tokens - but is this issue general across agents? Do agents know how much they'll spend?
Introducing Budget-Aware Agents (BAGEN): We study budget awareness across 4 envs & 5 frontier agents, and find structured failures in most of them.
👇
23
41
332
615,888
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!
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 the world won't pause;
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
8
438
Zihan "Zenus" Wang retweeted
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!
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 the world won't pause;
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
28
1,415
Zihan "Zenus" Wang retweeted
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 the world won't pause;
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>
10
50
273
22,204
Zihan "Zenus" Wang retweeted
Pausing the world while a model thinks leaves latency unresolved. In StarCraft, your base could be under attack while the model is still thinking. Glad to see Sooool exploring this. Hope future frontier models can truly play games in real time!
Excited to share a new repo: Sooool!
Sol itself playing games, frame by frame. No code-as-a-policy!
The environment changes while Sol processes the images. When an action arrives, it may no longer be correct. We built a research loop that accounts for this delay:
- Run up to eight model requests in parallel. Sol keeps looking at new frames and choosing actions while its earlier decisions are still in flight.
- Profile the target API and simulate its delays using cheaper inference during development. We only simulate the delay itself: Sol still sees the frames and chooses what to do at each step.
- Give Sol latency mean, std, P90 and P95. When it looks at a frame, it can plan for where things will be when its action arrives.
- Run episodes, inspect failures, update Sol’s instructions, and let it play again. Sol decides what lessons can enter later prompts.
We expect to spend about $50 per task on paid API calls, primarily for final verification rather than development. Most development runs use cheaper inference with simulated latency.
Touhou ↓
Sol chooses movement, shooting, focus and bombs from frames.
Findings from Sol in the auto-research loop:
- Predict where the boss will be, not where she was.
- Check the move AND the waiting position.
- More damage is not enough if it costs more hits..
Repo, experiment details, and the reproduction guide below. ↓
1
4
471
Sol plays Ping-Pong frame-by-frame!
Excited to share a new repo: Sooool!
Sol itself playing games, frame by frame. No code-as-a-policy!
The environment changes while Sol processes the images. When an action arrives, it may no longer be correct. We built a research loop that accounts for this delay:
- Run up to eight model requests in parallel. Sol keeps looking at new frames and choosing actions while its earlier decisions are still in flight.
- Profile the target API and simulate its delays using cheaper inference during development. We only simulate the delay itself: Sol still sees the frames and chooses what to do at each step.
- Give Sol latency mean, std, P90 and P95. When it looks at a frame, it can plan for where things will be when its action arrives.
- Run episodes, inspect failures, update Sol’s instructions, and let it play again. Sol decides what lessons can enter later prompts.
We expect to spend about $50 per task on paid API calls, primarily for final verification rather than development. Most development runs use cheaper inference with simulated latency.
Touhou ↓
Sol chooses movement, shooting, focus and bombs from frames.
Findings from Sol in the auto-research loop:
- Predict where the boss will be, not where she was.
- Check the move AND the waiting position.
- More damage is not enough if it costs more hits..
Repo, experiment details, and the reproduction guide below. ↓
4
22
1,944
Zihan "Zenus" Wang retweeted
Excited to share Sooool, our early exploration of coding agents tackling latency-sensitive tasks in real time. Its research loop tests ideas, learns from failures, and refines policies under simulated API delays before evaluating them against the real clock.
Excited to share a new repo: Sooool!
Sol itself playing games, frame by frame. No code-as-a-policy!
The environment changes while Sol processes the images. When an action arrives, it may no longer be correct. We built a research loop that accounts for this delay:
- Run up to eight model requests in parallel. Sol keeps looking at new frames and choosing actions while its earlier decisions are still in flight.
- Profile the target API and simulate its delays using cheaper inference during development. We only simulate the delay itself: Sol still sees the frames and chooses what to do at each step.
- Give Sol latency mean, std, P90 and P95. When it looks at a frame, it can plan for where things will be when its action arrives.
- Run episodes, inspect failures, update Sol’s instructions, and let it play again. Sol decides what lessons can enter later prompts.
We expect to spend about $50 per task on paid API calls, primarily for final verification rather than development. Most development runs use cheaper inference with simulated latency.
Touhou ↓
Sol chooses movement, shooting, focus and bombs from frames.
Findings from Sol in the auto-research loop:
- Predict where the boss will be, not where she was.
- Check the move AND the waiting position.
- More damage is not enough if it costs more hits..
Repo, experiment details, and the reproduction guide below. ↓
1
1
5
491
Zihan "Zenus" Wang retweeted
latency is such a pain for LLM game agents. most setups just pause the game and predict a few actions ahead. running requests in parallel is a simple, effective fix, love it!
abother view: if the agent picks up a world model of the game early on, it could roll forward from the stale frame and act on where things will be, not where they were 👀
Excited to share a new repo: Sooool!
Sol itself playing games, frame by frame. No code-as-a-policy!
The environment changes while Sol processes the images. When an action arrives, it may no longer be correct. We built a research loop that accounts for this delay:
- Run up to eight model requests in parallel. Sol keeps looking at new frames and choosing actions while its earlier decisions are still in flight.
- Profile the target API and simulate its delays using cheaper inference during development. We only simulate the delay itself: Sol still sees the frames and chooses what to do at each step.
- Give Sol latency mean, std, P90 and P95. When it looks at a frame, it can plan for where things will be when its action arrives.
- Run episodes, inspect failures, update Sol’s instructions, and let it play again. Sol decides what lessons can enter later prompts.
We expect to spend about $50 per task on paid API calls, primarily for final verification rather than development. Most development runs use cheaper inference with simulated latency.
Touhou ↓
Sol chooses movement, shooting, focus and bombs from frames.
Findings from Sol in the auto-research loop:
- Predict where the boss will be, not where she was.
- Check the move AND the waiting position.
- More damage is not enough if it costs more hits..
Repo, experiment details, and the reproduction guide below. ↓
4
1
10
748
Check out this one. really live-played by GPT-Sol!
It was 9️⃣ playing the game in the main thread
Excited to share a new repo: Sooool!
Sol itself playing games, frame by frame. No code-as-a-policy!
The environment changes while Sol processes the images. When an action arrives, it may no longer be correct. We built a research loop that accounts for this delay:
- Run up to eight model requests in parallel. Sol keeps looking at new frames and choosing actions while its earlier decisions are still in flight.
- Profile the target API and simulate its delays using cheaper inference during development. We only simulate the delay itself: Sol still sees the frames and chooses what to do at each step.
- Give Sol latency mean, std, P90 and P95. When it looks at a frame, it can plan for where things will be when its action arrives.
- Run episodes, inspect failures, update Sol’s instructions, and let it play again. Sol decides what lessons can enter later prompts.
We expect to spend about $50 per task on paid API calls, primarily for final verification rather than development. Most development runs use cheaper inference with simulated latency.
Touhou ↓
Sol chooses movement, shooting, focus and bombs from frames.
Findings from Sol in the auto-research loop:
- Predict where the boss will be, not where she was.
- Check the move AND the waiting position.
- More damage is not enough if it costs more hits..
Repo, experiment details, and the reproduction guide below. ↓
1
19
919
Excited to share a new repo: Sooool!
Sol itself playing games, frame by frame. No code-as-a-policy!
The environment changes while Sol processes the images. When an action arrives, it may no longer be correct. We built a research loop that accounts for this delay:
- Run up to eight model requests in parallel. Sol keeps looking at new frames and choosing actions while its earlier decisions are still in flight.
- Profile the target API and simulate its delays using cheaper inference during development. We only simulate the delay itself: Sol still sees the frames and chooses what to do at each step.
- Give Sol latency mean, std, P90 and P95. When it looks at a frame, it can plan for where things will be when its action arrives.
- Run episodes, inspect failures, update Sol’s instructions, and let it play again. Sol decides what lessons can enter later prompts.
We expect to spend about $50 per task on paid API calls, primarily for final verification rather than development. Most development runs use cheaper inference with simulated latency.
Touhou ↓
Sol chooses movement, shooting, focus and bombs from frames.
Findings from Sol in the auto-research loop:
- Predict where the boss will be, not where she was.
- Check the move AND the waiting position.
- More damage is not enough if it costs more hits..
Repo, experiment details, and the reproduction guide below. ↓
4
7
52
9,398
Sol plays Asterix from screenshots, directly choosing movement commands. Each direction stays held until its next reply arrives in real time.
Through autoresearch, it learns to:
- Prioritize escape over chasing food.
- Finish a reachable pickup before changing targets.
- After a pickup, choose the next direction instead of repeating the held action.
1
1
5
291
More examples in the repo: github.com/sooool-project/so…
The fun part isn’t just watching Sol play these games. Watching a failure turn into a lesson for the next run reminds me of learning to play them as a kid.
Give Sol a new environment. See what it learns!
5
264
Can gpt-sol do real-time robotics tasks that require latency awareness?
Yes.
We have a paper coming out soon :)
prediction: “big models are too slow for robotics” is about to age badly.
within weeks, expect many DeepSeek V4 Flash–scale JEVs: real-time, surprisingly close to Astra via distillation.
people underestimate how fast prefill can get.
11
2
138
8,909
Zihan "Zenus" Wang retweeted
Representing the world as editable, inspectable, and verifiable code has taken us remarkably far — from ARC-AGI-3 to 3D spatial reasoning!
Excited to release the full technical report and code for Schema!
Super proud of my first-year PhD student Guanning and the whole team! 🚀
On July 15, we released Schema, the first agent harness to reach ~99% RHAE on the ARC-AGI-3 public set.
Today, we’re sharing the full paper and code, with new results and a closer look at why Schema works.
4
26
1,694
🚨 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 transitions p(s′ | s, a), 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
19
3,054
Zihan "Zenus" 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
4
20
9,300