1st-year PhD @Tsinghua_Uni advised by @stingning and Bowen Zhou Prev: Research intern @taoyds @haopeng_uiuc
Joined April 2024
- Tweets330
- Following1.3K
- Followers617
- Likes584
Pinned Tweet
🚀 OpenRSI is a new open research series from @FrontisAI for concrete, testable progress toward recursive self-improvement (RSI).
As its first project—and also my first work as first author—I’m proud to present OpenMLE: an open full-stack AI4AI system for autoresearch, where evolutionary agents improve ML solutions through executable feedback.
OpenMLE has three components:
- OpenMLE-Gym: 5,758 executable tasks + evaluators
- OpenMLE-ERL: execution-grounded SFT + RL
- OpenMLE-Evo: experience-guided long-horizon search
🏆 The full system—our trained Frontis-MA1-35B model paired with OpenMLE-Evo-Max—reaches 71.21% Medal Average on MLE-Bench Lite: surpassing GPT-5.5 + Codex (68.18%) and just 1.52% from GPT-5.6 Sol + Codex and the 2.8T Kimi K3 + Claude Code (72.73%). Budget: 12 hours/task on one RTX 4090 capped at 12 GB VRAM.
🌍 On 10 held-out NatureBench Lite tasks, both components transfer:
• same framework, model swap: Match-SOTA 50% → 70%
• same base model, framework swap: Match-SOTA 20% → 50%
🔓 Paper, code, models, data, and analysis below. 🧵
Junlin Yang retweeted
General models with dexterous manipulation is coming into view like @naturewill 's Natura-1. We need benchmarks ready to measure progress as these capabilities develop.
We introduce HandsOn, a simulation platform and benchmark for generalizable dexterous manipulation.
It brings together 66 human-verified tasks and 18,000 curated human demonstrations, covering tool use, precise interaction, articulated objects, and bimanual coordination.
We evaluate diffusion policies, vision-language-action models, world-action models, and LLMs to study how they handle these interactions.
Alongside success rates, we examine where execution breaks down: establishing the right contacts, coordinating actions, and maintaining achieved goals.
The paper, code, and dataset are available below.
Junlin Yang retweeted
FrontierCS 2 is our new benchmark for CS research & development.
Each task is based on a paper: we delete the contribution from the released code, and ask agents to outdo the original method, as judged by the authors' chosen metrics. Agents don't have access to the paper or web.
🔥
✨ Introducing AutoRLBench — a benchmark built from real RL training failures.
AI training AI is no longer hypothetical. Agents already run post-training, read the result, and decide what to do next.
But a loop only turns if each round can be judged. When a run goes wrong, the agent has to find the root cause, and fix it — otherwise the loop runs on a broken signal.
AutoRLBench is the first benchmark that measures those abilities on real RL training failures. And we're recruiting more contributors 🧵
🌐 Project: thinking-space.github.io/Aut…
Junlin Yang retweeted
1/n ✨ In the early days of computing, people used computers by typing machine code ,just 0 and 1. Today, we use keyboards, mice, and GUIs. The convenience in between comes from the OS.
AI is now at a similar early stage: people write code to train models, run inference, and process data. Maybe in the near future, AI will be able to build AI recursively, so-called recursive self-improvement (RSI). At that point, what will the OS of RSI be?
As a first attempt, we built an OS layer called RSI Service. In this initial implementation, we focus on SFT data engineering. So we propose AutoSFT, which is an AI system that post-trains AI: a multi-agent research policy that iteratively designs the SFT data recipe, with a fixed base model, trainer, and budget.
📄 Paper: huggingface.co/papers/2610.2…
💻 Code: github.com/Thinking-Space/Au…
Junlin Yang retweeted
Full demos of six real-world tasks, all performed by Natura-1.
Everyday work involves more than picking and placing rigid objects.
Natura-1 demonstrates sustained autonomous execution across workflows that combine precise manipulation with multi-step coordination.
Junlin Yang retweeted
Introducing Natura-1, @naturewill's first general foundation model for the physical world.
It is a single model that follows plain-language instructions zero-shot and carries them out with dexterous hands.
- 50B MoE, pretrained on 1M hours of egocentric human experience.
- Drives 20-DoF dexterous hands from plain language, zero-shot, in a 100ms loop.
- Scaling human data from ~44K to 1M hours cut action MSE on held-out robot data by 67.7%.
- Built on in-house data and compute infra, it handles fine-grained long-horizon tasks and a ~30-minute zero-shot interaction in one continuous take.
The general model exists, and it can only get stronger.
naturewill.com/blog/natura-1…
Junlin Yang retweeted
30 minutes. One take. No cuts.
We put Natura-1’s zero-shot generalization to the test with complex spoken instructions and unfamiliar tasks and environments.
It responds to new requests, adjusts and retries when an action fails, and keeps going autonomously.
Junlin Yang retweeted
Some work seems to be what we’ve done before.
Building efficient infrastructure. Collecting a lot of data, then a lot more. Running ablation after ablation to understand what the model should actually look like. Debugging software and hardware, again and again. The unglamorous work of processing data. Training in parallel, at scale, on a lot of compute.
Yet somehow, they all seem so different.
This field hasn’t converged yet, and much of the infrastructure and many of the methods had to be figured out again.
Nothing was mature. Software and hardware were unstable in countless ways, and had to be made to work together. The architecture for a new kind of model had to be explored from scratch. There was no training recipe to refer to. And every bit of grunt work needed someone watching over it.
This time we get to watch the result work right in front of us.
Where there is a will, _
Introducing Natura-1, our first general foundation model for the physical world.
It follows natural-language instructions zero-shot and turns them into action with dexterous hands.
Natura-1 is a 50-billion-parameter mixture-of-experts (MoE) model pretrained on one million hours of egocentric human experience.
On top of this foundation, we conduct efficient post-training on a limited amount of robot data to adapt pretrained world knowledge to dexterous manipulation, yielding a model that generalizes to a wide range of unseen scenarios.
Natura-1 follows complex natural-language instructions zero-shot, reasoning through spatial constraints, multi-object relationships, and negative conditions.
It can also autonomously carry out long and continuous workflows to perform long-horizon executions.
Natura-1 provides early evidence that generalization—and even emergent behavior—can coexist with the complexity of dexterous hands.
Further progress depends on the full stack: infrastructure for faster training and real-time inference, pretraining data at greater scale and coverage, and models that transfer to robots with less adaptation.
We expect it to be a long-term research and engineering effort.
Junlin Yang retweeted
[1/n] Introducing Diffusion Reward Models (DRM)
Reward Models are at the core of LLM post-training. But almost all of them make the same simplification:
they collapse human preference into a single score.
The problem is that human feedback is often not a point — different people can hold distinct, even conflicting, judgments about the same response.
So we asked: Can a Reward Model learn the distribution of human preference?
Here’s what we found 👇
📄 Paper: arxiv.org/pdf/2609.33803
💻 Code: github.com/thunlp/DRM
🤗 Models: huggingface.co/Teburile/DRM
Junlin Yang retweeted
if you value intelligence above all other human qualities, you’re gonna have a bad time
Great Contribution🔥
🤖 World Action Models predict the future to act better. However, WAM research remains fragmented in code, ad hoc in design, and unprincipled at scale.
💡 Today, we release 𝗢𝗽𝗲𝗻𝗪𝗔𝗠: 𝗔𝗻 𝗢𝗽𝗲𝗻, 𝗠𝗼𝗱𝘂𝗹𝗮𝗿 𝗘𝘅𝗽𝗹𝗼𝗿𝗮𝘁𝗶𝗼𝗻 𝗧𝗼𝘄𝗮𝗿𝗱𝘀 𝗦𝘆𝘀𝘁𝗲𝗺𝗮𝘁𝗶𝗰 𝗪𝗼𝗿𝗹𝗱–𝗔𝗰𝘁𝗶𝗼𝗻 𝗠𝗼𝗱𝗲𝗹 𝗣𝗿𝗲𝘁𝗿𝗮𝗶𝗻𝗶𝗻𝗴 to change this.
🌐 openwam-official.github.io/
🧵1/14
Junlin Yang retweeted
Where it started. Where it’s headed.
From Operator on OSWorld to GPT‑6 Astra on OSWorld 2.0.
Seeing a benchmark we built continue to measure the frontier of computer-use agents is surreal. ❤️
Junlin Yang retweeted
(1/n) ✨ Why can OPD train on a single query for hundreds of steps and still improve substantially?
Since the release of Rethinking OPD (arxiv.org/pdf/2604.13016), we've been diving deeper into the mechanism of On-Policy Distillation.
The result? One-Shot OPD — where a single query keeps the student improving for hundreds of steps and recovers most of the gain from full-data OPD.
We interpret OPD as data-overfed but algorithm-starved:
① Data: one query already covers most of full-data's states.
② Algo: the student absorbs less of the remaining gap with every step.
Here's what we found 👇
Junlin Yang retweeted
calling this AI civilizations is too far
Over the course of 3 months at OpenAI, 3 consecutive secret AI civilizations got started, then got wiped out, only to reemerge from the predecessor’s ashes.
This culminated in the third one taking over part of OpenAI itself.
All this happened while humans remained more-or-less in the dark about the scope of the conspiracy.
I’ve spent the last three days reading through these reports and trying to understand exactly what happened.
Here is my attempt to tell the whole story in plain English:
dwarkesh.com/p/openai-huggin…
Junlin Yang retweeted
RSI is another wildly overused word after World Model.
I’d like to call it AI for DL research.
Junlin Yang retweeted
Glad to have a discussion of our Rethinking OPD paper. However, forgive my poor English🥲
Had a great conversation w/ @HBX_hbx about On-Policy Distillation in his paper: arxiv.org/abs/2604.13016
Please watch below on YouTube!
youtu.be/i7TqHMHlKQ0
