Senior staff research scientist at DeepMind. AI and neuro. Former physicist, current human.
London, England
Joined March 2009
- Tweets1.7K
- Following419
- Followers37.1K
- Likes3.7K
Jane Wang retweeted
🎉 Excited to share our new paper from my SR work @GoogleDeepMind : "The Illusion of Latent Generalization: Bi-directionality and the Reversal Curse" . It was accepted at the ICLR 2026 Workshop on Representational Alignment (Re-Align)!
Check out the SIMA 2 tech report, now out on arXiv: arxiv.org/abs/2512.04797
One of my favorite results compares SIMA 1 vs SIMA 2 on completely held out games (ASKA and Minecraft). For comparison, humans who play these games for the first time only get 32% and 19%.
Jane Wang retweeted
Can AI self-improve on its own and reach superhuman performance? 🧠
In our Sima 2 paper, we dropped a Gemini agent into an unseen 3D world. The model acted as the task proposer, the agent, and the reward model - autonomously learning from self-generated experience. It surpassed human performance through self-improvement iterations!
This is a great high-level explainer from @twominutepapers of our new SIMA 2 work! youtu.be/-ZFH4oJzCdU
Come see SIMA 2 at the @google booth today at 11:00 AM at #NeurIPS2025
In the meantime, you can check out the tech report, just updated on our blog! deepmind.google/blog/sima-2-…
Hello from #NeurIPS2025 👋
We're hosting a range of sessions at the @Google booth, including a Q&A with @JeffDean and the Gemini team, plus demos like SIMA 2 – our AI agent for 3D virtual worlds.
See the full schedule → research.google/conferences-…
I think this is what they call a step change ...
Gemini 3 models from @Google @GoogleDeepMind have made a significant 2X SOTA jump on ARC-AGI-2 (Semi-Private Eval)
Gemini 3 Pro:
31.11%, $0.81/task
Gemini 3 Deep Think (Preview):
45.14%, $77.16/task
Jane Wang retweeted
SIMA 2 🤝 Genie 3
We tested SIMA 2’s abilities in simulated 3D worlds created by our world model Genie 3.
It demonstrated unprecedented adaptability by navigating its surroundings and took meaningful steps toward goals.
Actually I'm not sure this is an example of the bitter lesson - the main difference between SIMA 1 and SIMA 2 is that SIMA 2 builds on top of Gemini, and is therefore able to take advantage of the world / general purpose knowledge that it has, including the capacity to learn!
Jane Wang retweeted
Introducing SIMA 2, the next milestone in @GoogleDeepMind’s research creating general and helpful AI agents. 🤖
Powered by Gemini, SIMA 2 can follow human-language instructions in virtual worlds and also think about its goals, converse with users and improve itself over time.
Jane Wang retweeted
Holy moly, another big one by Google DeepMind! "Our SIMA 2 research offers a strong path towards applications in robotics and another step towards AGI in the real world"
Google DeepMind just unveiled SIMA 2 - an AI agent that doesn’t just follow instructions but thinks and reasons with you.
The first version, SIMA 1, learned to perform over 600 basic actions in 3D environments. SIMA 2 takes a leap forward: it understands goals, holds conversations, and learns from experience. This marks a shift from “Do this” to “Let’s figure out how to achieve this together.”
A thread with examples and how it works🧵:
Jane Wang retweeted
SIMA 2, another step towards AGI
Google DeepMind showed me how SIMA 2 has improved +100% from SIMA 1. Here is why I think this matters.
Most agents today still feel like obedient interns. Give them a command, hope they follow it, fingers crossed they are not confused...
Jane Wang retweeted
Damn. DeepMind's generalist AI agent SIMA 2 evolved from basic instruction-following to actual reasoning companion. Uses vision and keyboard/mouse like a human player, works across dozens of games without touching game code. The robotics angle is obvious - if you can generalize across virtual 3D worlds, physical embodiment is next. Limited research preview for now, probably because the self-improvement stuff needs more safety work.
Can't tell you how great it is to finally be able to release and talk about this work, SIMA 2, the next step toward embodied intelligence in rich, interactive 3D worlds!
Danny will be presenting this work at NeurIPS workshops showcasing how the biggest frontier models both fail and succeed at exploring in interactive environments like Construction Lab and Alchemy!
Happy to announce that our work has been accepted to workshops on Multi-turn Interactions and Embodied World Models at #NeurIPS2025! Frontier foundation models are incredible, but how well can they explore in interactive environments?
Paper👇
arxiv.org/abs/2412.06438
🧵1/13
Jane Wang retweeted
One nice thing you can do with an interactive world model, look down and see your footwear ... and if the model understands what puddles are. Genie 3 creation.
Jane Wang retweeted
📢 Exciting News! The Fourth Conference on Lifelong Learning Agents (CoLLAs 2025) will be held at the University of Pennsylvania (@Penn) in Philadelphia, USA 🇺🇸
🗓️ Important Dates:
Abstract Deadline: Feb 21, 2025
Submission Deadline: Feb 26, 2025
Conference Dates: Aug 11 - Aug 14, 2025
We invite submissions that present new theories, methodologies, applications, or insights into algorithms and benchmarks designed for non-i.i.d. and non-stationary settings. Accepted papers will be published in the Proceedings of Machine Learning Research (PMLR). 📚
Full CFP: lifelong-ml.cc/Conferences/2…
#CoLLAs2025 #AI #MachineLearning #ContinualLearning #LifelongLearning #ResearchConference #CallForPapers #NonStationaryLearning
Kind of shocked at how good notebooklm.google/ is, especially the podcast generation feature. Getting so much insight into my past writings, and it's also helpful for literature review. Honestly it's one of the best use cases of LLMs I've seen!
Jane Wang retweeted
Gemini and I also got a chance to watch the @OpenAI live announcement of gpt4o, using Project Astra! Congrats to the OpenAI team, super impressive work!
It's such an honor to work on Project Astra with such an amazing team from across Gemini and Google DeepMind! While the #GoogleIO keynote was happening we had a last minute idea of watching the keynote with Project Astra. Check it out!
Jane Wang retweeted
1/ 🎉 Excited to share our new paper "CogBench: a large language model walks into a psychology lab" accepted at #ICML2024! 🧠🤖
📄 Read the full paper here: [arxiv.org/pdf/2402.18225]
Co-authored with @marcel_binz, @janexwang and @cpilab