Jiale Cheng

PHD student in CoAI Group, Tsinghua University, interested in LLMs and VLMs.

[Jiale Cheng @chengjiale4]

Excited to share what we’ve been building — Hyra-1.0, our first step toward AI systems that can improve research and engineering solutions through iteration. 🚀

[Tencent Hy @TencentHunyuan]

Introducing Hyra-1.0, the first version of Hunyuan Research Agent. 💡💡💡

Built to recursively improve solutions for performance-driven research and engineering tasks.

Explore our demos in AI4AI, AI4Science, and AI4Fun:
https://t.co/GClJMuT8XN

[Jiale Cheng @chengjiale4]

@IntCyberDigest https://t.co/qo6DoSugM5, maybe you are looking for Glyph lol!

[Jiale Cheng @chengjiale4]

@OpenBMB Congrats! Any plans to open-source ForgeTrain or share some lessons from building it?

[Jiale Cheng @chengjiale4]

Here’s a quick demo of the generate-from-scratch flow:

[Jiale Cheng @chengjiale4]

Love Claude Design, hate daily limits? 🛑

Meet Open CoDesign — the open-source, desktop-native AI design and web dev workspace.

🚀 Use the world’s most advanced models: GPT-5.4, Claude Opus 4.7, GLM-5.1, Gemini-3.1, or your own local models.

BYOK. No vendor lock-in. No usage-limit popups.
It’s Claude Design, but fully open. 🔓

🔗 Star on GitHub: https://t.co/KUxBuszXts

#OpenSource #AI #WebDev #BuildInPublic

[Jiale Cheng @chengjiale4]

Come and have a try!

[Z.ai @Zai_org]

Introducing GLM-5V-Turbo: Vision Coding Model

- Native Multimodal Coding: Natively understands multimodal inputs including images, videos, design drafts, and document layouts.
- Balanced Visual and Programming Capabilities: Achieves leading performance across core benchmarks for multimodal coding, tool use, and GUI Agents.
- Deep Adaptation for Claude Code and Claw Scenarios: Works in deep synergy with Agents like Claude Code and OpenClaw.

Try it now: https://t.co/WCqWT0qCQb
API: https://t.co/xDy1O6ZPcz
Coding Plan trial applications: https://t.co/qCM6cri0KK

[Jiale Cheng @chengjiale4]

@yan5xu @xDinoDeer 太牛了！怎么做到的，求经验🫡

[Jiale Cheng @chengjiale4]

What you see in the video:
1️⃣ Autonomous Deep Research: Using Tavily to scan 2024-2025 papers.
2️⃣ Code Execution: Writing and running its own evaluation harness.
3️⃣ Analysis: Generating plots and Markdown reports.

This is why we call it the AI R&D operating layer.
🔗 Project: https://t.co/Wa5NxZVEQi

[Jiale Cheng @chengjiale4]

Can an AI discover better prompts than a human? 🤖👨‍🔬

I asked PhD-Zero to investigate the optimal prompting strategies for Qwen-1.7B.

4 hours of autonomous research later, it didn't just find tricks—it produced a 10-page report with counter-intuitive insights that I hadn't even considered.

🎥 Watch the full end-to-end R&D loop:

#LLMs #autoresearch

[Jiale Cheng @chengjiale4]

The most fascinating finding? "Thinking Compression." 🧠⚡

While we often default to long Chain-of-Thought (CoT), PhD-Zero proved that for 1.7B models, concise reasoning actually outperforms verbose paths.

In small models, excessive "thinking" introduces logical drift. It literally out-researched me on the "physics" of its own kind. 🤯

[Jiale Cheng @chengjiale4]

The goal of PhD-Zero is to build the AI operating layer for R&D—equipping agents with specialized research skills instead of just writing code.

90% autonomous workload, but 100% human-directed. Check out the project and our vision here: 🔗 https://t.co/Wa5NxZV70K
The era of Autonomous AI R&D is here. 🚀

[Jiale Cheng @chengjiale4]

What happens when an AI Agent leads the R&D?
0% → 20.0% on AIME25 in just 48 hours. 🧠🚀

I deployed PhD-Zero (powered by Codex) as a 24/7 AI research intern to optimize Qwen-1.7B-Base.

11 iterations. 90% hands-off R&D, requiring only human confirmation at key milestones. Zero sleep, just high-speed execution. 🧵 on how this intern "thought" its way through:

cc @karpathy @bindureddy @Alibaba_Qwen

[Jiale Cheng @chengjiale4]

2/ The "Aha!" moment: Step 5. 🛠️

The agent autonomously identified a loss_mask mismatch that I would overlooked.

After the fix, supervised length jumped from 688 to 6552 tokens. The model finally "clicked" and started learning the actual reasoning. This self-healing is what separates an Agent from a script.

[Jiale Cheng @chengjiale4]

@karpathy Love the loop! I’m building PhD-Zero to give Claude Code/Codex a "PhD Skill-set."

It turns them into 24/7 autonomous researchers by standardizing the R&D workflow.

Let the agents handle the 2 a.m. experiments. 🧠⚡

🔗 https://t.co/KEQBLTgisM

[Jiale Cheng @chengjiale4]

@dingyi Try our work! (worked on both Win/macOS)

Repo: https://t.co/pv6EL0udGD

[Jiale Cheng @chengjiale4]

@coopernicus01 @claudeai May be you will be interested in our work (worked on both Win/macOS)

Repo: https://t.co/pv6EL0tFR5

[Jiale Cheng @chengjiale4]

@koltregaskes @claudeai May be you will be interested in our work (worked on both Win/macOS)

Repo: https://t.co/pv6EL0tFR5

[Jiale Cheng @chengjiale4]

Case3: Generate XLSX Spreadsheets 📉

[Jiale Cheng @chengjiale4]

What if Claude Cowork became truly open — and worked on both Win/macOS? 🧰🚀

Meet Open Cowork: an open-source Agent SDK desktop app with sandbox permissions, skills for PPTX/DOCX/Excel, and traceable execution.

Repo: https://t.co/pv6EL0udGD

[Jiale Cheng @chengjiale4]

Case2: Generate PPT from Files 📊