Guangzhao He

Student Researcher @GoogleDeepMind.
CS PhDing @Cornell.
BS with honors @ZJU_China.
I’m making AGI. I also make EDM.
Opinions are my own.

[Guangzhao He @AlexHe00880585]

@JiaweiYang118 huge congrats Jiawei!

[Guangzhao He @AlexHe00880585]

@yunongliu1 Congrats Yunong!

[Guangzhao He @AlexHe00880585]

@wenhaocha1 Congrats Wenhao! super exciting

[OpenRSI Foundation @OpenRSI]

As AI systems enter a recursive self-improvement loop, the central question is whether it can systematically move beyond human-designed methods to genuinely extend the scientific and intelligence frontier.

What’s missing is a neutral, open standard for evaluating these capabilities in real, production-scale intelligence development.

Today, we’re releasing OpenRSI-Index v0.1: evaluating whether AI can recursively improve itself and push the boundaries of intelligence and science, on production-scale clusters with 1k GPUs.

We turn fully open-source projects into autoresearch environments, with agent trajectories lasting 60+ hours. Building v0.1 took 100K+ H100-hours.

We’re building an ecosystem with and for the research community: let RSI benefit everyone, and let everyone shape RSI together.

We invite task contributors and compute partners to build this open benchmark with us - all contributors will be included as paper authors.
Shape RSI with us:
🌐 Website: https://t.co/unaB2yfQy4
🛠️ GitHub: https://t.co/vaRyltgbra
🤝 Contribute: https://t.co/PffRAhoIht

[Fuli Luo @_LuoFuli]

MiMo-V3 is getting a new architecture. The core of it, HySparse2, is out today.

Less prefill, a smaller KV cache, better long-context retrieval—and we got all three at once.

Compared with MiMo-V2.6's Hybrid SWA architecture:
• 5.02× lower prefill FLOPs at 1M tokens
• 4.5× smaller KV cache at 1M tokens
• Better MRCRv2 and RULER-v2 scores, plus lower AgentPPL and LongPPL

Why build a new architecture?
Agentic inference is a very different workload. Each round, a short action can return a long observation that needs to be prefilled, while the context keeps growing. That puts prefill cost, KV-cache size, and retrieval accuracy on the critical path at the same time.

HySparse2 tackles all three with two levels of KV sharing:
• KV Bridging: Following YOCO, full-attention layers in the cross-decoder build their K/V from self-decoder hidden states.
• KV Reuse: Within each hybrid block, sparse layers reuse the preceding full-attention layer's KV cache and selection indices.

Two more changes: token-level selection replaces block-level selection, and a forced window of recent tokens replaces the separate SWA branch, so local and global tokens share one KV cache. Since all cross-decoder KV caches now come from the self-decoder, prefill can stop once the self-decoder finishes.

Paper: https://t.co/REeEdd5pL7

[Guangzhao He @AlexHe00880585]

@wenhaocha1 woah is it that good!

[Guangzhao He @AlexHe00880585]

@devteamdrew @claudeai waaaaaaaat

[Guangzhao He @AlexHe00880585]

@aliteracy 5th floor is great :)

[Sam Padilla @theSamPadilla]

After 2+ years in the robotics data space, we are shutting @Eidon_AI down.

The thesis was right. But the business is brutally hard.

We close this chapter by open-sourcing everything we built and sharing lessons for anyone venturing into the space. https://t.co/pyL9IkjQPT

[Ben Poole @poolio]

sketch-to-simulation with Opus 5.5

[Furong Huang @furongh]

There’s a reason I stayed up all night on my birthday writing that article (https://t.co/fmD2nGlW10): I’ll be serving as an ICML 2027 Program Chair, alongside Weijie Su, Andrew Wilson, Julien Mairal, and Courtney Paquette. As someone who has worked on AI detection and will help run the conference, I feel a real responsibility to think more deeply about “AI slop” and what we should actually do about it.

With reports of 60K+ abstract submissions to ICLR 2027 (https://t.co/PUtya9TZen), how do we scale peer review without lowering the bar or burning people out? Recruiting enough reviewers is one challenge. Matching papers to the right expertise is another.

Here’s where I stand.

Using AI to improve your writing should not be punished. This matters especially for non-native English speakers. If AI contributed meaningfully to your ideas, experimental design, or analysis, be transparent about it. Either way, you are responsible for understanding, checking, and standing behind the work.

But **we should not ask reviewers to spend hours on a paper its own authors haven’t read, checked, or understood**. That standard should apply whether AI was involved or not. Submitting a paper is also a request for someone else’s time.

And “just run it through an AI detector” is not the answer. Research has documented false positives, including for non-native English writing, and ways to evade detection. A detector score is not a verdict on scientific quality. (https://t.co/U6WOgirdZD) Nor should confidential submissions be uploaded to external services without authorization and appropriate privacy safeguards. (https://t.co/GYZdJcdxkE)

I think desk rejection deserves serious consideration, but it needs to be based on demonstrable problems, not “this sounds AI-written.” What can we reliably identify before full review? Who makes those calls, and how do we catch mistakes? We need to protect reviewer time without screening out unconventional work just because it is unfamiliar.

**I’d love concrete suggestions on reviewer matching, early triage, and author accountability.** What have other venues tried? What worked and what backfired? How do we discourage careless submissions without making it harder for newcomers and less-resourced researchers to participate?

These are my personal starting views, not an announcement of ICML policy. We’ll do our best to serve the community, and your input would genuinely help.

#ICML2027

[Rujun Han @HanRujun]

Recursive Self-Improvement (RSI) is transforming how AI agents evolve, but have we considered that it needs regularization too?

🚀 Excited to share our latest work: RRSI (Regularized Recursive Self-Improvement of Agent Harnesses) https://t.co/Spy7EqHwwB!

[Guangzhao He @AlexHe00880585]

super cool results!

[HU Wenbo @wbhu_cuhk]

Introducing GAE🥳🥳

Generate once. Decode RGB + point clouds.

GAE generates directly in a geometry-native latent space. The same generated latent supports both appearance and 3D geometry.

The key choice is where generation happens. GAE compresses geometry-foundation features into a shared latent space. A conditional flow model generates directly in that space; RGB and geometry are decoded from its final state.

The video pairs show camera-conditioned world generation; the character pairs show text-to-image generation. These native showcase clouds are not reconstructed from generated RGB.

Code: https://t.co/euE8mZKI4G
Page: https://t.co/M0dMYwIOPp

[Guangzhao He @AlexHe00880585]

@Zichen2501 @LuoRundong0122 bitter lesson indeed!

[Guangzhao He @AlexHe00880585]

We built our 3D coding harness earlier this year (https://t.co/ZFSMjKfjny), back when none of the models we tested could really hit the bar.

Revisiting it now with stronger models like Astra, we’re seeing a sharp threshold: below it, the harness helps a lot. above it, almost not at all. This is mostly the same across all model providers.

A small warning for everyone building “agents”: all of today’s scaffolding will eventually get eaten by the base model.

Also, we're not done playing with this. Sharing more interesting stuff soon :)

[Yanjie Ze @ZeYanjie]

The Bitterest Lesson:
doing the right task > data > compute > algorithms

from https://t.co/WBNYfmtkQd

[Guangzhao He @AlexHe00880585]

I stand by my point now that iclr opened submission count reaches 55k

[Guangzhao He @AlexHe00880585]

cs academia is dying. let's please replace conferences with blog posts

[Zhenjun Zhao @zhenjun_zhao]

Can 4D Foundation Models Remember?

@AlexHe00880585, @ElorHadar, @weichiuma

tl;dr: benchmark->visual memory in 4D foundation models

https://t.co/cfdOmtQaBY

[s1r1us @S1r1u5_]

On July 25, we hacked OpenAI.

Two bugs let us take over ChatGPT/Codex accounts of OpenAI employees (+some unaffiliated users) and reach connected services: Outlook, Slack, GitHub, etc.

We proved it with a PR in OpenAI’s internal codebase . It took us <72h. 🧵

[Brett Adcock @adcock_brett]

The holy grail for robotics is being able to generalize: doing work in unseen places

We rented 30 homes in the Bay Area and are doing tasks without any new training