Jiani Wang

Senior undergrad @Tsinghua_IIIS. Intern at Impossible Research.
🎓 Applying for PhD, Fall '27

[Jiani Wang @Jiani_Wang_]

To figure out the mechanism of an unfamiliar environment, an agent has to work out what the world is and how it changes. That's what our [schema] does.

On ARC-AGI-3 Public, [schema] thinks like a physicist: it grounds the state and discovers the mechanism in one editable program. This induced world program makes action efficient: once it discovers a mechanism, it plans inside the model for free. And we achieve an amazing 99% RHAE.

Starting from here, mechanism discovery in much richer environments is what's next. Stay tuned!

[Haven Feng ✈️ CoRL @HavenFeng]

Today, we’re introducing [schema]: a harness reaching 99% RHAE with Opus 4.8 + Fable 5 and 95.35% with GPT-5.6 Sol on ARC-AGI-3 Public set.

[schema] makes an LLM think like a physicist. 🧵

[Haodong Wen @herrywen1]

(1/9)Suppose you're training an LLM on web data with limited math data. Should you place all your math data at the end of training, or mix it with web data earlier? The answer flips with the learning rate (LR) schedule!

At fixed compute and mid-training (MT) token budgets, earlier mixing gives lower MT loss when LR decays throughout training. With constant LR, placeing all MT data for the end works better.

We then vary when LR decay begins. The best tested strategy: keep LR high after MT data enters, then decay near the end.

See our new paper: "Data Schedules and Learning Rate Schedules Should Be Designed Together" !

[Andrea Zanette @Zanette_ai]

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! 🚀

[Haven Feng ✈️ CoRL @HavenFeng]

Schema is now fully released — code and paper are out! 🚀

Beyond ARC-AGI-3, Schema also tackles MazeBench and DIG (Discovery in Games).

The goal has always been broader: building agents that can discover, understand, and reason about how worlds work.

More in Guanning’s post 👇

[Guanning Zeng @guanningzeng]

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.

https://t.co/mj4f8cm4uz

[Yen-Jen Wang ✈️ CoRL @wangyenjen]

Practice makes perfect. A truly agentic robot shouldn’t just know what to do. It should also figure out how to get better.

What if a robot could watch a demonstration, figure out what it needs to practice, and learn how to get better through practice?

Introducing RPG: Guided Self-Improvement for Embodied Agents, a real-to-sim-to-real pipeline that turns real-world demonstrations into targeted practice.

RPG reconstructs practice tasks from demonstrations, uses demo videos to guide skill improvement, and continually updates a shared skill system from failures.

On real robots, RPG achieves 30/30 successes across 3 tasks. In simulation, it reaches 95.0% success across 22 manipulation tasks after 15 rounds of self-improvement.

Demo → Reconstruct → Guided Practice → Self-Improve → Go Real

[Haozhe Jiang @erichzjiang]

The used key is always bright.

Just released an agentic robot project RPG, where self-improving embodied agents play a heavy role. Wrote a blog on what motivated the project, what we learned, and some vision about what the future is like.

https://t.co/ykUJrCJf5d.

[Andrea Zanette @Zanette_ai]

Extremely proud of my students @amuseddaman and Xuecheng, and collaborator @g_k_swamy, for receiving an Outstanding Paper Award at COLM!

Congratulations to the team! 🎉

[Alvin Zhang @ZAlvin39105]

Existing continual learning methods still suffer from catastrophic forgetting. Our recent work found a simple recipe to improve 𝗹𝗼𝗻𝗴-𝗵𝗼𝗿𝗶𝘇𝗼𝗻 𝗺𝗲𝗺𝗼𝗿𝗶𝘇𝗮𝘁𝗶𝗼𝗻 substantially:

Composing multiple continual learning mechanisms extends memory lifetime and raises average final retention from 1.2% to 34.9% across 3 datasets: a 28-fold improvement over naive sequential SFT.

Project page: https://t.co/AeCJ5mxG6U
Alphaxiv: https://t.co/TzbwLOTqes
Arxiv: https://t.co/InHFQNFYuP
Code + datasets: https://t.co/PLthYR9RdO

[Runhan Huang ✈️ CoRL 2026 @RunhanH]

Excited to share our work accepted to CoRL 2026 @corl_conf !

We present SkillX: Unified Multi-Skill Policy Learning for Humanoid Soccer — a unified policy that enables humanoid robots to learn and compose diverse soccer skills, achieving smooth skill transitions and long-horizon multi-skill behaviors.

Project page: [SkillX project page](https://t.co/NPQfzMr1Hg)
Paper: [SkillX: Unified Multi-Skill Policy Learning for Humanoid Soccer (arXiv)](https://t.co/dALz24jeqL)

Video below 👇

#CoRL2026 #RobotLearning #Robotics #AI #EmbodiedAI #HumanoidRobots #Soccer

[Jiani Wang @Jiani_Wang_]

@HaochengXiUCB Amazing work!

[Haocheng Xi @HaochengXiUCB]

Open-source video generation is now faster than playback without compromising quality.

Introducing Video Delta Net (VDN): hybrid attention for live text-to-video with near-lossless quality.

VDN accelerates Minimax-H3 by 75 - 90 x, generating 14 seconds of 768p video in 11 seconds on 8× NVIDIA B200 GPUs.

Checkpoints + training/inference code + Technical Blog ⬇️
(1/6)

[Kairong Luo @openhonor]

How much would it cost you to pretrain a 2B LLM from scratch? $1M? $100K?

🚀Announcing Puro-2B, with a fully open recipe.

You can train a model that matches Qwen2-1.5B on RTX 5090s for less than $5090!

$4.4K → beats Qwen2-1.5B
$6.9K → approaches Qwen2.5-1.5B

Check here👇

[Kaifeng Lyu @vfleaking]

Excited to introduce Puro-2B, along with its fully open recipe that lets you train a Qwen2-1.5B-level LLM on RTX 5090 GPUs for less than $5090!

As someone who has always been curious about why LLMs work so well, one of the biggest challenges today is that we not only lack a good theory, but are also losing the ability to run controlled training experiments at scale, systematically vary key factors, and gather the empirical evidence needed to understand how LLMs learn from data.

Training LLMs has simply become too expensive.

There are now a few great open or almost open training recipes for models with billions of parameters, but reproducing them can still cost hundreds of thousands of dollars. As this makes many basic questions about LLMs inaccessible to academic labs, we really hope to bring this cost down far enough that academic labs can train and study genuinely capable models from scratch.

Puro is a full-stack engineering effort to reduce training cost, which mixes engineering and research across data, optimization, numerical precision, hardware, and training strategy. It also builds on some of our own previous research:

📈 Curriculum Model Averaging (CMA): data curriculum designs have to work together with moderate learning rate decay and model averaging (https://t.co/dkufW4qMbz)
◉ Hyperball / MuonH: an optimizer wrapper that constrains weight and update norms, and directly controls the effective (angular) step sizes for stable and efficient training (https://t.co/MZEkev4Ts7)

There is still a long way to go, and we would be very happy to hear your feedback and ideas!

[Alvin Zhang @ZAlvin39105]

1/N Finally found time to put out my second blog post, based on small experiments I ran in my spare time.

If you splice an INCORRECT chain of thought into a model and let it keep going, can it still get there? 🧐 Yes, right up until the wrong reasoning is finished.

https://t.co/ToFZ3audZs

[Haodong Wen @herrywen1]

Finally, our paper is on arXiv! 🎉

Huge thanks again to my amazing collaborators Yiran @zhangyir22, Yingfa @yingfachen, and Kaifeng @vfleaking! 🙌

We complete the global-minima analysis of 2D-RoPE and prove that an $\ell_\infty$-constrained global minimizer found by one-layer 2D-RoPE can learn copy with length generalization.

More broadly, we hope this sheds light on how positional encodings shape a model’s learning behavior beyond this specific setting. ✨

arXiv: https://t.co/Mc0wBang7d

P.S. Always happy to chat about theory! 🤩

[Jiani Wang @Jiani_Wang_]

Huge thanks to the whole team: @guanningzeng, @wenjie_ma, @shaofeng_y27736, @ChenyangWa70207, @lustralisk95, @akanazawa, @wodenimoni, @xiuyu_l, @Zanette_ai and @HavenFeng!

[Jiani Wang @Jiani_Wang_]

To figure out the mechanism of an unfamiliar environment, an agent has to work out what the world is and how it changes. That's what our [schema] does.

On ARC-AGI-3 Public, [schema] thinks like a physicist: it grounds the state and discovers the mechanism in one editable program. This induced world program makes action efficient: once it discovers a mechanism, it plans inside the model for free. And we achieve an amazing 99% RHAE.

Starting from here, mechanism discovery in much richer environments is what's next. Stay tuned!

[Haven Feng ✈️ CoRL @HavenFeng]

Today, we’re introducing [schema]: a harness reaching 99% RHAE with Opus 4.8 + Fable 5 and 95.35% with GPT-5.6 Sol on ARC-AGI-3 Public set.

[schema] makes an LLM think like a physicist. 🧵

[Andrea Zanette @Zanette_ai]

ARC‑AGI‑3 is an RL problem: can agents discover the hidden mechanisms of an unfamiliar world through interaction, as efficiently as humans?

Today we introduce [Schema], a harness that enables frontier models to build, test, and refine executable world models. [Schema] reaches ~99% RHAE in our evaluation of the ARC‑AGI‑3 public set.

Led by my incoming PhD student @guanningzeng, together with an amazing team!

[Xiuyu Li @xiuyu_l]

Mountain: climbed. Time to explore the world.

Our [schema] harness achieves 99% on ARC-AGI-3 Public by thinking like a physicist with coding agents. The rate of improvement in capability is amazing, and a lot of tasks need to be redesigned from the foundations up. More to come!