Ruofeng Yang

ARIS (Auto Research In Sleep), 16k star in GitHub
Looking forward for Disssuison on AUTOresearch
Intern in Naive AI @naiveailab

[Ruofeng Yang @WShuiyin]

详细的流程和multi Agent协同在ARIS (Auto Reserch In Sleep,18k 🌟)里面完整的得以workflow以及mcp形式得以利用，大家有兴趣可以去康康

[Ruofeng Yang @WShuiyin]

大科研时代，研究生多Agent高效科研上岸指北
感觉多智能体协同一方面可以充分利用到各家模��的特点，另外一方面方便大家薅各���plan，会比较穷哥们儿友好
✅找idea
找idea推荐是在有明确的领域和自己简单的阅读自己的paper和感悟之后，可以写下来，然后以6 Astra为主结合Gemini 进行讨论然后大规模fanout idea先广搜再深搜筛选idea做预实验验证可靠性 (对应idea discovery workflow
Gemini 4对于找idea来说基本上是必不可少的，因为GPT系列的模型特别容易过度防御，然后找出来的idea往往是奇形怪状或者为了novel 而novel的(一眼就是coding过度对齐了
而Gemini 4虽然coding较弱，但是想idea 以及讨论非常好用

✅定实验plan (plan refiner
定实验plan需要比较高智能的模型来做最开始的plan，推荐6 Astra和op 5.5一���来讨论定，然后需要强调不要过度防御比如扫描超参数做表格啥的)可以用Hero选择(Hero-Anti-Overdefense来约束)

✅做实验刷bench (experiment bridge workflow
可以用经济高效但是幻觉率低的力工模型来做，比如6.1 sol, grok 4.7 或者一些比较实惠的模型

✅写paper (paper writing workflow
建议Gemini 4作为绝对主力来进行文章撰写，写出来有人味而且很灵动，然后codex 之类的做事实性校验(注意HERO选择禁止过度防御)然后搭配ARIS出图workflow 使用GPT Image 2.5 出图降低AI味道

[Ruofeng Yang @WShuiyin]

论原生家庭以及应试教育对模型行为的影响

OAI就是应试教育父母😮‍💨
Google是经典的快乐教育的父母(虽然孩子coding差点就差点吧，agy也是真烂👶
A/则是高素质克制但淡漠的父母的教育，疏远淡漠礼貌但克制且强大🧐
老马的grok则是纯黄毛父母来的🤪

Gemini 4完全感觉是一个enfp快乐小狗来的，每一句话先肯定加感叹号然后积极的回答你🥹

很明显没有过度对齐也没有像6Astra一样的过度自责和自卑，

感觉真的是非常有趣的感受，哈基米情绪价值确实给的足，孩子笨点就笨点，但是确实快而且很快乐，然后在等待Astra的时候有了很多乐趣

[Ruofeng Yang @WShuiyin]

大模型大理论Bench严肃发布 Gemini恐成赢家
我宣布在大理论时代Gemini 赢麻了
最近和友友严肃整理大理论Bench Ack各位友友
感觉有点好玩的
Gemini以3/3的成绩占据榜首，榜单要求一个大理论里面至少有三家模型，即可入选
#ai #大模型 #科研 #论文 #人工智能 #AI工具 #博士生日常 #互联网大厂 #AIGC

[Ruofeng Yang @WShuiyin]

@peter9863 Data interation towards real data manifold than Better solution？Sounds great idea

[Ruofeng Yang @WShuiyin]

大模型本质是大米，有感觉么
Mimo MiniMax Kimi Gemini
哈哈哈哈哈哈被朋友逗死了

[Ruofeng Yang @WShuiyin]

在论文不如纸的情况下，研究生应该如何上岸
如题，大家找好用的AutoResearch的项目用就好了
在AI洪流下，总要和自己和解和上岸
尽自己一切努力就好，好好休息就好
大家加油

[Ruofeng Yang @WShuiyin]

@MaxForAI 我感觉是不是真的会有一个产品类似于JoyAI Code来对标Claude code啥的hhhh

[Ruofeng Yang @WShuiyin]

EasyAIOffer开源|手机随时刷的秋招AI面经🎉
转眼就要金九银十的秋招时刻了，大家也进入了关键时期，ARIS-In-AI-Offer也500🌟了
我也进行了系统性的更新，规约到了一个手机上随时可以打开的html命名为🌟easyaioffer🌟然后更新了更多新的内容🎉🎉
使得大家在手机上也可以随时随地复习(虽然今年AI coding很多，但是基础知识还是要有的

秋招 AI 岗面试，一个链接刷完👇
直接search ARIS-in-AI-Offer就可以看到啦
或者easyaioffer💥github💥io 也可以
希望大家在秋招更加轻松

大家有啥希望改进的也可以提issue或者PR

💥涵盖了35 篇 cheat sheet，8 个方向：
📎基础（Attention/Transformer/归一化/优化器/分词）
📎后训练与推理（RLHF·DPO·GRPO / 推理模型 / KL）
📎架构与系统（MoE / KV cache / 量化 / 分布式 / 推理服务 / 预训练流水线）
📎生成模型理论（扩散 / Flow Matching / VAE）
📎生成系统（图像 / 视频 / 3D / 扩散后训练与蒸馏）
📎多模态、Agent、具身与世界模型

每篇都是：
✅ 公式推导 + 从零 PyTorch 代码（能跑）
✅ 25 道高频题，分 L1 必会 / L2 进阶 / L3 顶级 lab
✅ 中英双语，单文件 HTML，手机上直接看，不用装东西
一共 856 道题，都展现出来了

每篇都过了ARIS type的跨模型审稿，使用的是Fabel5.1和GPT6 协作完成，可以说是比较高质量了，很多文章我也会亲自审阅

[Ruofeng Yang @WShuiyin]

17k🌟自主科研ARIS|开源国模(Deepseek V4pro-preview + GLM 5.1)协作超越5.6Sol
转眼间ARIS(Auto Research in Sleep，https://t.co/nQyF5B7pIR)已经14000 star了🎉
最近可算抽出来时间把ARIS在各类AutoResearch 的Benchmark上面测了测😋😋
结果还是挺出乎我意料的，发现在已有的AutoResearch Bencmark上，ARIS使用DeepSeekV4 Pro和GLM5.1(这也是ARIS的平替建议组合)根据workflow进行对抗性配合和协作就能够轻松追平GPT5.6 sol🔥🔥
这说明对于比较长程Auto Research的任务来说，是不是任务执行结果的可验证和可回溯比纯粹的模型能力更重要？
只要配合的好，未必不能做到更好🧐🧐
(btw国模真的太强了，我都不敢想最先进的国模能��到什么境地，有时间测测国内各大玩家的模型配合的怎么样，感觉可以给我很多惊喜啊🤩🤩
希望可以和大家多交流！
ARIS也在持续更新，我们最近也更新了针对Theory Proof的workflow 和lean 验证 以及自己独立的ARIS-Code，欢迎大家给我提issue和讨论🙇🏻‍♂️🙇🏻‍♂️

[Ruofeng Yang @WShuiyin]

2026 年，视频生成真的还需要扩散模型么？
In 2026, do we really still need diffusion models to generate videos?

Agents, Blender, and careful planning can already take us surprisingly far.

A machine that only paints straight lines draws its first curve to spare a paper flower—and discovers another way to create, just as we explore a purely agentic approach to image and video creation, without diffusion models, aiming for more precise rendering and more faithful physics..

一辆只会画直线的小车，为了绕开一朵纸花，画出了第一道弯，也发现创造可以有另一条路。这就像我们尝试不依赖扩散模型，纯粹由 agent 驱动图像与视频创作，以面向更加准确的渲染和物理。

This 60-second short explores that same idea. Agents developed the story, wrote code for the scenes, animation, and music, then rendered, reviewed, and revised.

No diffusion models were called—even to generate reference images.

That’s what we’re exploring with ALIGN-video: give agents the tools and room to plan, build, observe, and improve their own work.

Do we still need diffusion models? Yes. The motion here still feels a little mechanical. But seeing how far an agent-driven workflow can go makes this path worth exploring.

https://t.co/abV3CAMUK9

[Ruofeng Yang @WShuiyin]

我发现opus 5.5 好喜欢说英文 就算我多次给他说要说中文 但是还是给我突然用英文
这是怎么回事(btw 他真的很好用

[Ruofeng Yang @WShuiyin]

还有更多比如复杂的p5.js 迭代几十个loop 画清明上河图啥的以及ARIS审稿，感觉也蛮不错的，我整理整理也放出来，感觉还蛮有趣的

[Ruofeng Yang @WShuiyin]

AI4AI4AI开始叠罗汉了hh
用NaiveAI的Naive-N0.5-Flash 尝试了训练AlphaZero，他自主运行15h后轻松打过baseline了，基本上我就给了一个commnets就睡觉去了
可以看html 看具体的展示 https://t.co/q6oShgeaMp

[Ruofeng Yang @WShuiyin]

刚刚intern的Naive AI 发布了首个模型Naive-N0.5-Flash, 位列开源模型前列🎉🎉
309B MoE, native 1M context, up to 2,000 tok/s.
看Benchmark包括我自己实测都是都是挺能打的
Naive AI 还用N0.5-Flash 达到了一个World Arena的SoTA（多模态人默默哭泣了一下，之后都不用自己训练WM了
SWE Terminal Bench 以及 Sol Exec，NanoChat的Bench得以展示 还是在国模第一梯队的
而且模型还是很轻量而且MIT开源的

[NaiveAI @naiveailab]

Introducing Naive-N0.5-Flash: Building Frontier AI with AI

🔹 309B MoE, 15.5B active: top-tier in coding, leading in AI R&D.
🔹 Native 1M context, no full-attention layers (hybrid SWA + DSA).
🔹 Inference runtime built by AI: up to 2,000 tok/s in Ultrafast mode.

Weights are open today under MIT license.

🔗 Tech blog: https://t.co/P1e8pBeJJM
🤗 Hugging Face: https://t.co/v2YvKfuguc
💻 GitHub: https://t.co/QjJ7cjFlkL
🌐 https://t.co/sIc7nZJH1L

[Ruofeng Yang @WShuiyin]

Diffusion is Die? 💥ALIGN: Agentic Loop Image GeneratioN Without Diffusion

After trying Fable 5.1 and GPT-6 Astra, it really feels like we’re entering a golden age of coding agents. GPT-6’s 3D generation is getting to a point where I can actually use it, and drawing with pure code and p5.js is becoming surprisingly capable too.

The images here use no diffusion or autoregressive image generation model. A coding agent writes the program that draws them.

But long, complex scenes still give these agents trouble. They tend to produce repeated building blocks, similar houses and mechanical details. Getting a complete image is one thing; getting all those little relationships to work is much harder.

So I built ALIGN : Agentic Loop Image GeneratioN

The main demo is a Wuhan version of *Qingming Along the River*. I wanted to see whether the agent could carry the reference’s craft into a different city, beyond simply reproducing the original scene.

Even the workflow diagram in Figure 1 is drawn entirely in p5.js. The code, iteration history and design decisions are open source, covering the Wuhan scroll, *A Thousand Li of Rivers and Mountains*, and ALIGN’s own method figure.

The two ideas I most want to share are Loop and Adversarial Review

1️⃣ Loop: keep thinking after rendering.

The first draft had plenty of problems: repeated houses, floating figures and weak spatial relationships. Each round follows the same process: render, inspect the image, edit the code, then render again.

Sometimes changing a parameter helps. Sometimes the drawing method needs a rewrite. Sometimes an edit makes things worse and gets reverted. The Wuhan scroll went through 11 versions; the evolution images and score chart show how it developed.

2️⃣ Adversarial Review: get a reviewer who will actually challenge the result.

Claude writes the code. Codex examines the image: where the occlusion is wrong, where things look copied and pasted, and where the drawing falls short of the reference.

A reviewer from another model family brings a different perspective, and its criticism becomes concrete work for the next iteration. This is something ALIGN carries over from my ARIS project. In the comparison shown, the Claude + Codex run produced a visibly stronger result than the Codex + Codex run.

I built this to explore how far coding agents can go. Replacing diffusion isn’t the goal. Even in the final version, small details in the people and buildings still look mechanical. That’s where I think diffusion could be useful as a further rendering and refinement step.

There’s also an interesting visual parallel: our p5.js process starts with linework and builds up detail in stages. That progression, along with the repeated loop, reminds me of diffusion. I’ll explore this more in the upcoming blog post.

Code, skills for Claude Code and Codex, and the examples:
https://t.co/abV3CAMUK9

#GenerativeArt #CodingAgents #p5js #AI #AIGC #GPT #Claude #OpenAI

[Ruofeng Yang @WShuiyin]

ARIS-Movie-Director 🎬
Based on ARIS (https://t.co/kBWOGiafGk 12k star, welcome to star!)

Long-horizon AI image/video generation has two failure modes nobody really fixes:
📎Drift — identities, settings, earlier decisions quietly forget themselves
📎Linear self-critique — every frame is judged by the model that drew it; errors compound, no one independent checks

Sound familiar? Those are the *same* failure modes as long-horizon autoresearch agents. So I took the two features that fix it in ARIS and pointed them at generation — agent does the orchestration, diffusion just renders:

📚 research-wiki: persistent, *auditable* memory: every frame logs where it came from, how many rollbacks, why, and the repair prompt.
🗣️cross-model adversarial review: multi-agent debate across model families, blind token-diff, single-vote veto; *any* agent can trigger a rollback, even across frames.

From a vague concept it authors an asset library → outline → storyboard → comic.json,
then runs an audited spiral per panel: draw → blind cross-model gate (**looks right ≠ passes**) → KEEP / RETRY → write audit trace → assemble → release (PNG + single-file HTML viewer).

I had it make ARIS's own promo comic (https://t.co/VcUla9jrb7): a 19-scene / 24-frame pixel-art movie, through 100+ audited rollbacks, until every agent agreed the on-screen text, the logic, and every character identity held. Full wiki open-sourced.

Long-horizon image → done. Video (open models) → in progress.
MIT · open-source 👇
https://t.co/5jqhYWBLRJ

#AIGC #GenerativeAI #LLM #Agents #Diffusion #ai