Staff Research Scientist @ Google DeepMind. Visiting Lecturer @ Harvard University. Previously CS PhD @ MIT CSAIL. Opinions my own. https://xitter.cf/t.co/UJRqylNAsN
Cambridge, MA
Joined September 2013
- Tweets135
- Following203
- Followers1.9K
- Likes402
Our team @GoogleDeepMind is hiring 3 PhD Student Researchers for 2027 to work on exciting, cutting-edge projects focused on understanding, evaluating, and aligning frontier AI agents and multi-agent systems. Signal your interest here forms.gle/2onGFBnyk7MaCdJn8 and spread the word!
Zi Wang, Ph.D. retweeted
Introducing Gemini 3.8 Flash, another upgrade in under a month! And the new 3.8 Flash Cyber pushes the frontier of cyber defense. Relentless progress 🚀 More on the models & our cyber security work in the blog - happy building! blog.google/innovation-and-a…
Thrilled to share our latest work! Do LLMs actually know what to ask when they don't have all the facts? Spoiler: they usually stop asking way too early. Check out the details in the thread below! 👇 Congrats @_yepeng and team for the release!
New preprint! Do LLMs know what to ask, and when?
Website: zitniklab.hms.harvard.edu/MT…
Paper: arxiv.org/abs/2608.14808
(1/n)
Zi Wang, Ph.D. retweeted
New preprint! Do LLMs know what to ask, and when?
Website: zitniklab.hms.harvard.edu/MT…
Paper: arxiv.org/abs/2608.14808
(1/n)
Zi Wang, Ph.D. retweeted
When an LLM is underperforming, we tweak the prompt. Why not do the same for video models solving visual reasoning tasks?
Visual prompt engineering changes the start frame appearance but not the task, improving performance.
📄 arxiv.org/pdf/2607.25537
🌐 visual-prompt-engineering.gi…
Zi Wang, Ph.D. retweeted
We believe AI can be a dedicated research partner to help discover the next breakthrough.
Enter Co-Scientist: our latest Gemini-based multi-agent system that can generate, debate and evolve novel hypotheses for complex scientific problems 🧵
Zi Wang, Ph.D. retweeted
We’ve released TECCI, a new dataset focusing on challenging image edits! (Including visual puzzles and much more)
Like DOCCI, all images are donated to the public domain, so you get the bytes with the data. See Aishwarya’s thread for more details and pointers.
Releasing a new image editing benchmark -- TECCI: Tricky Edits of Collected and Curated Images.
Paper: arxiv.org/abs/2606.01213
Project website: google-deepmind.github.io/te…
Check out Proactive Co-Creator on
@GoogleAIStudio , a human-AI belief alignment demo I vibe coded: aistudio.google.com/apps/bun…
🧠 See & edit the AI's uncertainty via belief graph. It asks clarifying questions before creating!
📷 Try Image ➔ Story ➔ Video. You can even remix it!
Music by Lyria 3. Research Showcase: zi-wang.com/human-ai/
Relevant paper:
✨ Proactive Agents & Belief Alignment arxiv.org/abs/2412.06771 with @MeeraHahn, Wenjun Zeng, @NithishKannen, Rich Galt, @BadolaKartikeya , @_beenkim
We've extended the belief graph to an even cooler idea: Atomic Concept Edits. Check out arxiv.org/abs/2602.00092 joint work with @NehaKalibhat @anonymani0 @_beenkim et al!
Zi Wang, Ph.D. retweeted
A decade ago in Korea, AlphaGo showed AI’s potential.
Together with the Korean government, we’re now looking at how this technology can help accelerate scientific discovery and create new opportunities for economic growth across the region. 🇰🇷
Find out more → goo.gle/4wbqUbJ
I'll present our ACE paper @aistats_conf #Morocco this Sunday, on behalf of @NehaKalibhat @anonymani0 @prasNLP @_beenkim et al.
ACE has been a super useful tool for us to explore prompts and gain critical insights on model behavior. Code open-sourced! github.com/google-deepmind/a…
Thrilled to share that our paper on "Interpreting and Controlling Model Behavior via Constitutions for Atomic Concept Edits" has been accepted at AISTATS 2026! 🚀🚀
Read more about how input mutations can be mapped to interpretable behavioral insights.
arxiv.org/abs/2602.00092
🧵
Zi Wang, Ph.D. retweeted
Evaluation shouldn't block the progress of AGI development.
We’re open-sourcing ProEval, a new framework to scale GenAI development without the massive price tag.
💰100x cost reduction
🎯 <1% error margin
🚀 Read the full @GoogleDeepMind tech report here:
Evaluating GenAI is too slow and expensive.
Our new paper introduces ProEval, achieving 8-65x fewer samples needed for 1% accuracy!
It uses Bayesian quadrature to efficiently estimate model performance and catch failure cases early. More in arxiv.org/abs/2604.23099
Evaluating GenAI is too slow and expensive.
Our new paper introduces ProEval, achieving 8-65x fewer samples needed for 1% accuracy!
It uses Bayesian quadrature to efficiently estimate model performance and catch failure cases early. More in arxiv.org/abs/2604.23099
Check out our open-sourced code and data here! github.com/google-deepmind/p…
Joint work with my amzing collaborators Yizheng Huang, Wenjun Zeng and Aditi Kumaresan
Fresh off arXiv! 🚀 We’re using Atomic Concept Edits as powerful exploration strategies to build rulebooks for model behavior shifts when concepts are manipulated. A new framework for adversarial steering and failure discovery! arxiv.org/pdf/2602.00092
See you at AISTATS 26!
Thrilled to share that our paper on "Interpreting and Controlling Model Behavior via Constitutions for Atomic Concept Edits" has been accepted at AISTATS 2026! 🚀🚀
Read more about how input mutations can be mapped to interpretable behavioral insights.
arxiv.org/abs/2602.00092
🧵
