Ziche Liu

CS PhD student @ UCSD | CUHK-Shenzhen ’26
LLMs · Agents · World Models · AI for Science

[Rikiya Takehi @rikiyatakehi]

We asked GPT-6 Astra to edit Qwen3.5-0.8B weights to improve math.
It directly changed only 0.016% of weights, fixing 0.8B's reasoning loops & improving math up to 6x and even instruction following 3x, without gradients or training.

We call this Weight-Space Programming: frontier LLMs can now natively read/write neural network weights! 🧵
Blog: https://t.co/bOWCcFhCVq

[Nan Huang @NanHuang99]

Can AI agents really work like scientists?

We introduce DISCERN, a benchmark for automated research across:

🔎 Data integrity
📊 Analysis validity
💡 Hypothesis generation

A key component is counterfactual evaluation: can agents reason from evidence even when it conflicts with established expectations?

Across 203 tasks, 8 life-science tracks, and 8 models, perfect performance drops from 60.8% → 34.2% → 0.6% from L1 to L3.

DISCERN also profiles agents’ working styles, revealing where autonomy breaks down and where human–AI collaboration still matters.

🌐 https://t.co/r0EshIgLra
📄 https://t.co/CFumRCtShL

#AIforScience #HumanAI #AutomatedResearch #AIAgents #ScientificDiscovery

[Yue @Yue1650749]

What could you discover with AI agents on your team?
Join the Galileo Hackathon at UC San Diego HDSI
🧪 Track 1, Nov 4-18: design a scientific ML task
⚡ Track 2, Nov 21: solve fresh tasks with SIDERIUS
🏆 Awards in both tracks, GPU & API credits provided

Open to US universities & labs 👇
https://t.co/rGk9TP3D9u

[Ziche Liu @QubitChinchilla]

Excited to finally see RobotAgentBench out! Can robots figure out what to do, actually do it, and recover when things go wrong? That's what this benchmark tests in realistic everyday household settings.
Huge kudos to @Lingjun_Mao , @williamcao1943 , @HeLukun11667 , and the amazing team for pulling off such an ambitious project!

[Lingjun Mao @Lingjun_Mao]

🤖 Introducing 𝗥𝗼𝗯𝗼𝘁𝗔𝗴𝗲𝗻𝘁𝗕𝗲𝗻𝗰𝗵, a comprehensive benchmark for evaluating 𝗼𝗽𝗲𝗻-𝘄𝗼𝗿𝗹𝗱 𝗿𝗼𝗯𝗼𝘁 𝗮𝘂𝘁𝗼𝗻𝗼𝗺𝘆 in realistic household environments.

Robots can now perform an increasingly wide range of manipulation tasks. But how far are we from robot assistants that can operate autonomously in the 𝗼𝗽𝗲𝗻 𝘄𝗼𝗿𝗹𝗱?

RobotAgentBench evaluates 𝗩𝗟𝗔/𝗪𝗔𝗠 𝗽𝗼𝗹𝗶𝗰𝗶𝗲𝘀 𝗮𝗻𝗱 𝗱𝗶𝘃𝗲𝗿𝘀𝗲 𝗿𝗼𝗯𝗼𝘁 𝗮𝗴𝗲𝗻𝘁 𝗵𝗮𝗿𝗻𝗲𝘀𝘀𝗲𝘀 across 𝗳𝗶𝘃𝗲 𝗱𝗶𝗺𝗲𝗻𝘀𝗶𝗼𝗻𝘀: task discovery, context reasoning, manipulation, failure recovery, and robustness. 𝗙𝗼𝘂𝗿 𝗼𝗽𝗲𝗻-𝗲𝗻𝗱𝗲𝗱 𝗵𝗼𝘂𝘀𝗲𝗵𝗼𝗹𝗱 𝘀𝗰𝗲𝗻𝗮𝗿𝗶𝗼𝘀 test how these capabilities work together to fulfill high-level user needs.

What we find:
• 𝗠𝗮𝗻𝗶𝗽𝘂𝗹𝗮𝘁𝗶𝗼𝗻 ≠ 𝗮𝘂𝘁𝗼𝗻𝗼𝗺𝘆. Standalone VLA/WAM policies struggle with task discovery and open-ended scenarios. Agent harnesses substantially improve performance, but reliable open-world autonomy remains challenging.

• 𝗛𝘆𝗯𝗿𝗶𝗱 𝗖𝗼𝗱𝗲–𝗩𝗟𝗔/𝗪𝗔𝗠 𝗣𝗼𝗹𝗶𝗰𝘆 𝗽𝗲𝗿𝗳𝗼𝗿𝗺𝘀 𝗯𝗲𝘀𝘁. Combining learned policies with code-based control achieves the strongest performance among evaluated configurations.

• 𝗥𝗦𝗜 𝗳𝘂𝗿𝘁𝗵𝗲𝗿 𝗶𝗺𝗽𝗿𝗼𝘃𝗲𝘀 𝗿𝗼𝗯𝗼𝘁 𝗮𝗴𝗲𝗻𝘁𝘀. Refining reusable tools and accumulating experience lead to additional gains across iterations.

• 𝗕𝗲𝘀𝘁 𝗰𝗼𝘂𝗽𝗹𝗲 𝘂𝗻𝗱𝗲𝗿 𝗼𝘂𝗿 𝗱𝗲𝗳𝗮𝘂𝗹𝘁 𝗮𝗴𝗲𝗻𝘁 𝘀𝗲𝘁𝘁𝗶𝗻𝗴: 𝗔𝗻𝘁𝗶𝗴𝗿𝗮𝘃𝗶𝘁𝘆 𝗖𝗟𝗜 (𝗚𝗲𝗺𝗶𝗻𝗶 𝟰 𝗔𝗿𝗴𝗼𝗻) + 𝗜𝗻𝘁𝗲𝗿𝗻𝗩𝗟𝗔-𝗔𝟭.𝟱. This pairing performs best among the planner–executor combinations we evaluated, achieving an open-ended score of 10.4, compared with 8.4 for fully code-based Codex (GPT-6 Astra) and 1.1 for standalone π0.

👇 Project website + more details in the thread. …1/11

[Lingjun Mao @Lingjun_Mao]

🤖 Introducing 𝗥𝗼𝗯𝗼𝘁𝗔𝗴𝗲𝗻𝘁𝗕𝗲𝗻𝗰𝗵, a comprehensive benchmark for evaluating 𝗼𝗽𝗲𝗻-𝘄𝗼𝗿𝗹𝗱 𝗿𝗼𝗯𝗼𝘁 𝗮𝘂𝘁𝗼𝗻𝗼𝗺𝘆 in realistic household environments.

Robots can now perform an increasingly wide range of manipulation tasks. But how far are we from robot assistants that can operate autonomously in the 𝗼𝗽𝗲𝗻 𝘄𝗼𝗿𝗹𝗱?

RobotAgentBench evaluates 𝗩𝗟𝗔/𝗪𝗔𝗠 𝗽𝗼𝗹𝗶𝗰𝗶𝗲𝘀 𝗮𝗻𝗱 𝗱𝗶𝘃𝗲𝗿𝘀𝗲 𝗿𝗼𝗯𝗼𝘁 𝗮𝗴𝗲𝗻𝘁 𝗵𝗮𝗿𝗻𝗲𝘀𝘀𝗲𝘀 across 𝗳𝗶𝘃𝗲 𝗱𝗶𝗺𝗲𝗻𝘀𝗶𝗼𝗻𝘀: task discovery, context reasoning, manipulation, failure recovery, and robustness. 𝗙𝗼𝘂𝗿 𝗼𝗽𝗲𝗻-𝗲𝗻𝗱𝗲𝗱 𝗵𝗼𝘂𝘀𝗲𝗵𝗼𝗹𝗱 𝘀𝗰𝗲𝗻𝗮𝗿𝗶𝗼𝘀 test how these capabilities work together to fulfill high-level user needs.

What we find:
• 𝗠𝗮𝗻𝗶𝗽𝘂𝗹𝗮𝘁𝗶𝗼𝗻 ≠ 𝗮𝘂𝘁𝗼𝗻𝗼𝗺𝘆. Standalone VLA/WAM policies struggle with task discovery and open-ended scenarios. Agent harnesses substantially improve performance, but reliable open-world autonomy remains challenging.

• 𝗛𝘆𝗯𝗿𝗶𝗱 𝗖𝗼𝗱𝗲–𝗩𝗟𝗔/𝗪𝗔𝗠 𝗣𝗼𝗹𝗶𝗰𝘆 𝗽𝗲𝗿𝗳𝗼𝗿𝗺𝘀 𝗯𝗲𝘀𝘁. Combining learned policies with code-based control achieves the strongest performance among evaluated configurations.

• 𝗥𝗦𝗜 𝗳𝘂𝗿𝘁𝗵𝗲𝗿 𝗶𝗺𝗽𝗿𝗼𝘃𝗲𝘀 𝗿𝗼𝗯𝗼𝘁 𝗮𝗴𝗲𝗻𝘁𝘀. Refining reusable tools and accumulating experience lead to additional gains across iterations.

• 𝗕𝗲𝘀𝘁 𝗰𝗼𝘂𝗽𝗹𝗲 𝘂𝗻𝗱𝗲𝗿 𝗼𝘂𝗿 𝗱𝗲𝗳𝗮𝘂𝗹𝘁 𝗮𝗴𝗲𝗻𝘁 𝘀𝗲𝘁𝘁𝗶𝗻𝗴: 𝗔𝗻𝘁𝗶𝗴𝗿𝗮��𝗶𝘁𝘆 𝗖𝗟𝗜 (𝗚𝗲𝗺𝗶𝗻𝗶 𝟰 𝗔𝗿𝗴𝗼𝗻) + 𝗜𝗻𝘁𝗲𝗿𝗻𝗩𝗟𝗔-𝗔𝟭.𝟱. This pairing performs best among the planner–executor combinations we evaluated, achieving an open-ended score of 10.4, compared with 8.4 for fully code-based Codex (GPT-6 Astra) and 1.1 for standalone π0.

👇 Project website + more details in the thread. …1/11

[Haoze He @HaozeHe_Hector]

Try it on https://t.co/Xmjb9TddPw and to play against one of our best models!

We gave open-source decision models two games. It did not go well. 🫠

🎮 Red Alert 2: the base model sold its own base. Then lost.
💥 Tank War: v1 spun in circles. v2 found the fire button… and shot itself.

So we let them keep learning: play → train → upgrade the harness → repeat. A few rounds later: model can evolve without human intervention

Then we put different models in one arena and let them fight it out. So far, no human has beaten the trained models. 😅

The lesson: a decision model isn't finished when it's released. Real environments keep throwing new situations at it, and a frozen model can't adapt. It needs continuous learning!

It's all open source: how we train these models to play, end to end. Next, we're turning it into training infra for any decision model.

⭐ Star it, fork it, or bring your own model to the arena:
GitHub: https://t.co/oAPAOkBRn0
RA2: https://t.co/xMv6VNaE11
Tank War: https://t.co/xttz2fjeWe

I also appreciate all the suggestions and feedback from: @b_vasilescu , Greg Ganger, @p0nk , @lateinteraction , @matei_zaharia , and @heathercmiller

And thanks to all the Contributors to the demo and the github repo!: @Jayzou3773, @brocoliman38951,
@ShangyinT , @Xuan04263449 , chenyang yang, tianyu chen , @AlexZhang152005, @Nickzz12138,@JasonWuzh, @Leow_sss730, @ChenMoneyQ , xingyuan ding, @JunchengLi , @LakshyAAAgrawal , @ao_qu18465

[Ziche Liu @QubitChinchilla]

Joint work with @yeeelow233, Ziche Liu, @JunxiaCui1120, @JasonWuzh, Yiqian Wang, @Jayzou3773, @Lingjun_Mao, @NanHuang99, @NaichengYu, @KaijieZhu07, @YueMa137, @KunZhou23339193, @LetianPeng and @shangjingbo.

[Ziche Liu @QubitChinchilla]

Can LLMs discover scientific laws?

Not even GPT-6 can uncover half of our hidden laws!

• A good fit is not always a valid law.
• Models can recall known laws but struggle to discover new ones.
• Models can generate a good law but don't recognize it.

[Ziche Liu @QubitChinchilla]

SciLaws-Bench: 118 problems from 381 papers, tested in real and parallel worlds.

Paper: https://t.co/gHVueXCIis
Project page: https://t.co/MdSINnPC8L
Code: https://t.co/hqnqrBZAb7
Data: https://t.co/8Y1NGvmoaj