Junlin Yang รีทวีตแล้ว
Junlin Yang
230 posts
Junlin Yang
@junlin45300
Incoming PhD @Tsinghua_Uni advised by Bowen Zhou and @stingning Research intern @haopeng_uiuc @taoyds
เข้าร่วม Nisan 2024
1.1K กำลังติดตาม143 ผู้ติดตาม
OPD has been blowing up recently, yet many open questions still remain. What does it really take to make OPD work in practice? This work goes deep into the details — highly recommended and well worth a read!🔥
Bingxiang He@HBX_hbx
1/n ✨ Introducing our new work: Rethinking On-Policy Distillation of Large Language Models OPD is now a core technique in LLM post-training (Qwen3, MiMo, GLM-5...). But here's the uncomfortable truth: it often doesn't work. We systematically study the phenomenology, mechanism, and recipe of OPD. Here's what we discovered 👇
English
Towards unified digital agents! 👊
Once again, this shows that code has already become a powerful and natural interface for digital agents.
Shibo Hao@Ber18791531
🍫 CocoaBench v1.0 is out! CocoaBench is a benchmark for unified digital agents, built around open-world tasks that require composing 💻 coding, 👀 vision, 🌐 search. Since our first research preview last December, we have expanded the benchmark substantially with community contributed tasks, and spent months testing and refining the tasks, evaluations, and agent runs. Some takeaways: • Even the best agent system reaches only 45.1% on CocoaBench v1.0. • Coding agents like Codex are already surprisingly strong on general tasks beyond software engineering. • Stronger agents tend to push more of the work into code. • Open source models still lag behind leading frontier models on these general tasks. 👇More on the website and in the paper #AI #Agents #LLM #Benchmark #CocoaBench
English
Junlin Yang รีทวีตแล้ว
Introducing Project Glasswing: an urgent initiative to help secure the world’s most critical software.
It’s powered by our newest frontier model, Claude Mythos Preview, which can find software vulnerabilities better than all but the most skilled humans.
anthropic.com/glasswing
English
Junlin Yang รีทวีตแล้ว
Surviving scale is becoming a real problem. Will AI labs eat everything?
Does each domain still need its own answer or just dissolve in the scaling pot?
Anthropic@AnthropicAI
Introducing Project Glasswing: an urgent initiative to help secure the world’s most critical software. It’s powered by our newest frontier model, Claude Mythos Preview, which can find software vulnerabilities better than all but the most skilled humans. anthropic.com/glasswing
English
In 2026, coders stopped being synchronous executors and became **async schedulers**: less time doing every step themselves, more time organizing context, allocating attention, and validating outputs.
English
if you can reply to this, you're awesome and cool :)
English
Junlin Yang รีทวีตแล้ว
The weight of a yes is measured by the no’s before it.
WeChat@Weixin_WeChat
Today, we are officially opening the capability to integrate #OpenClaw into #Weixin. With the launch of the #WeixinClawBot, users can use Weixin as a dedicated messaging channel for OpenClaw. Now, you can send and receive messages with OpenClaw just like texting a friend. #AIAutomation #AI
English
Junlin Yang รีทวีตแล้ว
I (finally) put together a new LLM Architecture Gallery that collects the architecture figures all in one place!
sebastianraschka.com/llm-architectu…
English
Junlin Yang รีทวีตแล้ว
✨ [ICLR 2026] How Far Can Unsupervised RLVR Scale LLM Training?
The dream: models can improve themselves without human supervision.
The reality: sometimes they can only sharpen what they already believe.
Intrinsic rewards struggle to scale LLM training because they follow a rise-then-fall pattern that makes collapse mathematically inevitable.
But that's not the end of the story. We find unsupervised RLVR (URLVR) is particularly well-suited for test-time training and quantifying model priors.
The full picture 👇
📄 Paper: arxiv.org/abs/2603.08660
🧪 GitHub: github.com/PRIME-RL/TTRL
(1/n)
English
Agentic test-time scaling (TTS) is effective -- until you find its inherent limits.
💡We show that classic TTS methods offered limited practical gains due to two fundamental limitations: the context ceiling and the verification gap. 🧵
Check the website: general-agentbench.github.io
English
@stingning @OpenAI @openclaw @steipete Totally agree! It’s getting clear that systems/evals are backward-shaping how models evolve. As we enter AI's 'second half,' the power to define what a 'good problem' or 'good application' actually looks like is really important.
English
Junlin Yang รีทวีตแล้ว
Goat
Hao Zhang@haozhangml
Can’t believe I get to say this -- deeply honored to be named a 2026 Sloan Research Fellow: today.ucsd.edu/story/2026-slo… Early faculty life is… "hyper-intense": teaching, advising, hiring, papers, grants; and trying to build a lab culture you’ll still be proud of years later. There were many weeks where it felt like we were building the plane mid-flight, burning plenty of midnight oil along the way. Over the past few years, I’ve been incredibly lucky to work with amazing students and collaborators on a chain of OSS project: Vicuna → Chatbot Arena → vLLM → DistServe → LMGame → FastVideo; each one then pushed forward way further by people far beyond our lab. This award feels less like a finish line and more like fuel for the lab, for our students, and for the next set of systems we haven’t built yet. A core principle of us is building "open-source research that ships." At the same time, it’s hard not to feel a mix of excitement + uncertainty + anxiety about where CS is heading. Coding agents are improving so fast that I am feeling the AGI first-handedly. I have gone back to builder mode -- only more productive than ever -- outside of my faculty admin work. I’ve watched friends and colleagues hit numbers that would’ve sounded like science fiction a year ago (e.g., 100+ commits/day). So what does it mean to “do great computer science” when baseline productivity keeps jumping? For me, it makes “research that ships” more important, and even raises the bar. The leverage shifts toward taste and problem selection, principled system design, and translating ideas into reliable artifacts. We're excited to keep proving that through real systems people can use! Deeply grateful to: - My students and collaborators — for the ideas, execution, and drive. - @HDSIUCSD , Dean @GuptaUcsd, and my @UCSanDiego colleagues — for building an environment where ambitious work can happen. - @nvidia and @mbzuai (and other compute sponsors) — for support that helped us move faster and turn ideas into real artifacts. Even as the interface changes, the need for efficient compute and solid infrastructure only grows. Most of all: credit to the students at @haoailab. You’re the reason any of this is worth doing. Keep building and shipping!
English
Junlin Yang รีทวีตแล้ว
Junlin Yang รีทวีตแล้ว
🚨New paper: "What Makes LLM Agent Simulations Useful for Policy Practice?"
@simile_ai is making a great launch, but can these LLM agent simulations actually help real institutions make better decisions?
We spent a year working with policymakers to answer this simple question.
The answer is yes—but perhaps not how you'd expect.
👇 THREAD 👇
[Link to paper: arxiv.org/abs/2509.21868]
[1/n]
English
"Collective reasoning under distributed information" is a fascinating and under-explored topic. It represents a core capability for **real-world multi-agent collaboration** in open environments, and this work provides an excellent evaluation for it!
Yuxuan Li@YuxuanL_
We’ve expanded our benchmark into HiddenBench: a 65-task, theory-grounded, extensible benchmark for evaluating collective reasoning under distributed information. We tested 15 frontier models (🚨spoiler🚨: Gemini is the clear winner) and uncovered key bottlenecks in multi-agent LLMs coordination. Check out the full update on arXiv! 📄 Paper: arxiv.org/abs/2505.11556
English
Junlin Yang รีทวีตแล้ว
Everyone is talking about Self-Evolving 🤖. But here’s the hard question: How do we actually evaluate it?
Are models truly learning new skills, or just recalling?
🚀We propose a new benchmark SE-Bench, targeting a core primitive of evolution: Knowledge Internalization. 🧵
English
Junlin Yang รีทวีตแล้ว
Finally it's here!
Qwen@Alibaba_Qwen
🚀 Introducing Qwen3-Coder-Next, an open-weight LM built for coding agents & local development. What’s new: 🤖 Scaling agentic training: 800K verifiable tasks + executable envs 📈 Efficiency–Performance Tradeoff: achieves strong results on SWE-Bench Pro with 80B total params and 3B active ✨ Supports OpenClaw, Qwen Code, Claude Code, web dev, browser use, Cline, etc 🤗 Hugging Face: huggingface.co/collections/Qw… 🤖 ModelScope: modelscope.cn/collections/Qw… 📝 Blog: qwen.ai/blog?id=qwen3-… 📄 Tech report: github.com/QwenLM/Qwen3-C…
English