Hy4 by @TencentHunyuan is now officially the #1 best Chinese model in Agent Arena!
With over 12K+ real-world, long-horizon agentic sessions from our global community of users, Hy4 has also placed Tencent Hunyuan as the only Chinese lab in the top 5. It is now showing +5.4% net improvement, and sits firmly on the Agent Arena Pareto frontier at a $0.57 median cost per task.
Congrats again to the @TencentHunyuan team!
👋Hy4 is here, new with vision.
Firmly among the leading open models.
More productive everywhere: finance, frontend, game, officework, research…
Built responsibly, with cyber/biosecurity in mind.
Try it out and let us know!
🔗:https://t.co/ZzL4DtgH3h
Building good RL environments is hard.
We're introducing a way to turn anything into RL-environments with REA using the open source AgentEnv framework from @ScaleAILabs
Any site, app, or game can now easily be turned into an env 😉
🌍 Seeing agents build worlds with code, I keep coming back to one thought:
🎯 I’d make that bet again.
In June 2025, we released IR3D-Bench, my co-first-author work published at NeurIPS 2025 Datasets & Benchmarks.
We asked a simple question: if a model claims to understand a scene, can it reconstruct it?
Give it an image. Ask it to predict the objects’ positions, shapes, sizes and materials. Turn those structured scene parameters into a Blender render. Put the reconstruction next to the original.
Now there is something concrete to inspect. Did it miss an object? Put it in the wrong place? Get the spatial relations right?
A fluent answer can hide uncertainty. A misplaced object cannot.
Our experiments showed a gap worth paying attention to: models could recognize many attributes and use rendering tools, yet precise spatial reconstruction remained difficult. Feeding the rendered result back to the model helped improve some failed cases.
Recently, I’ve been following work from other teams exploring related questions:
🧱 3DCodeBench: procedural 3D assets from text or images.
🎬 4DCodeBench: reconstructing geometry, appearance and motion from video as graphics programs.
🧠 Code as Worlds: executable world hypotheses for physical reasoning.
These are different tasks, with distinct contributions from their respective authors. I’m excited to see the research direction develop; I’m not claiming that these projects originated from ours.
Here’s why I keep betting on code as world representation.
Once a model expresses a scene through an executable description, it has to commit to details. Where is the object? How large is it? What changes over time? Which relationships must hold?
Those choices can be run, inspected and challenged. Change a coordinate. Render another view. Check whether the result still agrees with the observation.
There are limits. A convincing image, correct motion and successful interaction are three different tests. IR3D-Bench studies static synthetic scenes; it does not solve full physical-world modeling.
But that is a concrete place to start, and a direction I want to keep working on.
An AI says it understands the world. Show me what it can reconstruct.
▶️ We’ve put together a 55-second demo of IR3D-Bench. Watch the video, compare the reconstructions, and judge the idea for yourself.
🌐 Project & demo: https://t.co/2GL5MXTiO4
📄 Paper: https://t.co/W3hdrSuuaw
💻 Code: https://t.co/c3o2vlZwQh
@razaaitech Thanks! Seeing the reconstruction next to the input makes the errors much easier to spot. We found that getting object attributes right and placing objects accurately are still quite different challenges.
Incredible work with all co-authors of RobotWorld!
🤖 From Computer Use to Robot Use: How far are general-purpose agents from acting in the physical world?
AI agents can write code, operate computers, and call tools. But what happens when those tools are robotic arms, mobile robots, or drones?
RobotWorld tests whether agents’ existing vision, coding, and reasoning capabilities can translate into reliable robot task execution.
🦾 84 tasks. 8 robot embodiments. 5 domains.
The benchmark spans manipulation, mobile manipulation, locomotion, driving, and flight. Agents observe images and robot states, control robots through tool interfaces, and write programs to support their calculations.
Task success is determined by the environment—not by the model declaring “done.”
📊 How do current models perform?
We evaluated 5 models on the same 84 tasks, yielding 420 scored trajectories:
• GPT-6 Astra: 16 tasks completed (19.0%)
• Claude Opus 5.5: 13 tasks (15.5%)
• Kimi K3: 2 tasks (2.4%)
• DeepSeek V4.1 Flash: 1 task (1.2%)
• Gemini 3.8 Flash: 1 task (1.2%)
These results reflect performance under our simulation tasks, interfaces, and evaluation budgets; they do not establish real-world deployment readiness.
🔍 Beyond the scores, the trajectories tell a richer story.
Agents write image-segmentation programs, estimate camera parameters through exploratory movements, and build approximate dynamics models to generate control commands.
Yet these capabilities do not consistently lead to success:
📍 The arm reaches the intended position, but the object is not securely grasped.
🔄 Feedback arrives, but the strategy does not improve.
✅ The goal remains unmet, but the agent believes the task is complete.
💡 The gap between “can see, calculate, and code” and “can complete a robot task” remains substantial. Reliable execution requires a closed loop connecting perception, computation, action, and verification.
RobotWorld provides tasks, execution traces, and explicit success criteria to help researchers identify where that loop breaks—and guide future model training and agent design.
🌐 Project page: https://t.co/5LLprgnUqW
📄 Paper: https://t.co/quUuKvJDDt
💻 Code: https://t.co/selLPckTWo
🤗 Hugging Face Papers: https://t.co/cHojWnTFA5
We welcome feedback, discussion and contribution! If you find the work useful, consider an upvote on HF Papers or a star on GitHub. Thanks for your support! 🙏
🎬 Below: the paper’s first page and a 45-second introduction, featuring agents in action. 👇
💍Ring Forcing: Towards Precise Long-Term Memory for Autoregressive Video Diffusion
Can a video model remember an object after it disappears—and bring back the same one a minute later? 👀
📄 Paper: https://t.co/GcaA088zuS
🌐 Project: https://t.co/RiDOJ9EdvP
#ECCV2026! 🎉
Introducing JarvisHub—a canvas-native agent harness for long-horizon multimodal creation.
✓ Canvas as memory
✓ Inspectable & recoverable
✓ Image, video, web & slides
From one-shot outputs to complete projects.
GitHub: https://t.co/ygu1oMxznT
Paper: https://t.co/3Q3MNGXFGv