Zhoujun (Jorge) Cheng

This account is verified. Learn more

@ChengZhoujun

RL scaling @thinkymachines | Ph.D. @UCSanDiego

San Diego blankcheng.github.io Joined November 2021

778 Following 1.4K Followers 465 Posts

Latest posts 5

Pinned post

Zhoujun (Jorge) Cheng @ChengZhoujun Jan 20

Pretraining has scaling laws to guide compute allocation. But for RL on LLMs, we lack a practical guide on how to spend compute wisely.

We show the optimal compute allocation in LLM RL scales predictably.

↓ Key takeaways below

Zhoujun (Jorge) Cheng @ChengZhoujun 6d

try inking 2 small!

Thinking Machines Verified account @thinkymachines 6d

Introducing Inkling 2 Small. Scoring 31 on the AA Intelligence Index, it's a big jump over its predecessor.

Inkling 2 Small is available for fine-tuning and inference on Tinker today, with full weight release planned soon.

thinkingmachines.ai/news/inkling-2…
View quoted post on X

Zhoujun (Jorge) Cheng @ChengZhoujun Jul 8

Our work on compute-optimal RL is at ICML 2026 this week! — IsoCompute Playbook: Optimally Scaling Sampling Compute for RL Training of LLMs

📍 Poster #1813 @ HALL A, Thu Jul 9, 2:30–4:15pm KST, Seoul. I can't make it this time, but my amazing co-author @TongtongLiang99 will be there — go say hi!

Zhoujun (Jorge) Cheng Verified account @ChengZhoujun Jan 20

Pretraining has scaling laws to guide compute allocation. But for RL on LLMs, we lack a practical guide on how to spend compute wisely.

We show the optimal compute allocation in LLM RL scales predictably.

↓ Key takeaways below

View quoted post on X

Zhoujun (Jorge) Cheng @ChengZhoujun Jun 26

I think good benchmarks test what your model can improve; great ones quietly point at where the value is and what's worth improving next. For me the tell: you instantly wonder why it didn't come out sooner, and you want to try it today.
OSWorld V2 grounded in such real long-horizon everyday use feels like that.

XLANG NLP Lab Verified account @XLangNLP Jun 26

Two years ago, we built OSWorld 1.0 — the benchmark that became the standard for computer-use agents. Agents now score 83.5% on it. Problem solved?

Not even close.

🚀Today we introduce OSWorld 2.0: Benchmarking Computer Use Agents on Long-Horizon Real-World Tasks.

What's new:
🎯 108 real-world workflows, each ~1.6 hours ⏱️ for a skilled human
⚙️ ~318 tool calls/task vs. ~30 in OSWorld 1.0
🌍 Grounded in authentic artifacts & stateful user profiles
⚡ Captures real phenomena: dynamic environments, streaming interaction, cross-source reasoning, implicit-state inference & more

📊 Best results: Claude Opus 4.8 reaches the highest accuracy at 20.6%, while GPT-5.5 is far more token-efficient but plateaus near 13%. No one is close to solving real computer use.

🏠 Homepage: osworld-v2.xlang.ai
📄 Paper: github.com/xlang-ai/OSWor…
💻 Code: github.com/xlang-ai/OSWor…
🤗 Dataset: huggingface.co/datasets/xlang…

🧵 [1/8]

View quoted post on X

Zhoujun (Jorge) Cheng @ChengZhoujun May 26

Very solid data work here on cua envs. Massive value in both the outcome and the execution!

Bowen Wang Verified account @BowenWangNLP May 26

RLVR has become the recipe for agentic post-training. But for Computer-Use Agents, the bottleneck is not the algorithm, it is the data. 🐌

🚀 We introduce CUA-Gym: a scalable, lightweight synthesis engine that turns arbitrary task queries into verifiable RLVR data for computer-use agents. The largest open CUA RLVR dataset to date:

🎯 32,122 verifiable RLVR tasks with programmatic setup scripts + rewards
🌐 110 environments: 16 desktop apps + 94 synthesized mock web apps
🏆 Qwen3.5-based CUA models trained with GSPO reach 72.6% on OSWorld-Verified and 56.6% on WebArena

📄 Paper: huggingface.co/papers/2605.25…
🏠 Homepage: cua-gym.xlang.ai
🤗 Dataset: huggingface.co/datasets/xlang…
💻 Codebase: github.com/xlang-ai/CUA-G…
🧩 Environments: github.com/xlang-ai/CUA-G…
🧵[1/6]

View quoted post on X