Interesting read! Take-away: interpretably tools (logit-lens, J-lens, steering methods) can probe how "physics-knowledge" and "relationships" are stored in LLMs.
The internal metaphysics of an LLM: An LLM represents physical relationships in a form we can both read and steer, and providing insights into post-training signals that help models develop deeper generalizations of physical principles. Across a large set of materials-science laws, the direction of a shift inside the model revealed whether increasing one quantity makes another rise, fall, or stay the same. That pattern persisted when we rearranged the equations and when we removed them and described the problem in words. By nudging the model’s internal activity, we could test which relationships actually helped produce its predictions. New work published in AI for Science (link and press release below)
To measure a spring's stiffness, you pull on it and see how far it stretches. We “pulled” on three versions of Google’s Gemma language model by reversing a single physical change in a question. Then we measured the shift in its hidden state, the set of numbers the model uses as it works toward an answer. Two mathematical lenses helped us read those numbers. The direct logit lens asks which words the model would favor if it had to answer at that stage. A logit is a score for a possible next word or word fragment. The Jacobian lens accounts for how later stages typically reshape that signal. That is, a Jacobian tells us how sensitive each number downstream is to a small change in one upstream. When we described grooves forming along the boundaries of tiny crystals in steel, this lens ranked corrosion first, although the question never named it, revealing some of the inner thoughts of the LLM and how it reasons over physical principles.
We also built a graph to see which problems the model treated as similar. Each dot represented a question, and we connected two dots when the patterns of numbers inside the model looked alike. Within a given mechanism, two questions could become neighbors simply because both described an increase. What does that increase mean for the material? More pores, or empty spaces, make a material less stiff. More stiff particles make it stiffer. To see whether the model captured that distinction, we asked each question twice, once with the input increasing and once with it decreasing. The direction of the shift inside the model revealed which way it expected the material to respond. Even without equations or step-by-step hints, both larger models captured the correct direction for all laws that predicted a change.
Then we pushed from the inside: A metal is made of tiny crystals called grains; in the metals we tested, smaller grains mean more boundaries between crystals. Those boundaries make it harder for the metal to change shape permanently, so its strength goes up. We kept each question exactly as written and made a small adjustment to the model’s internal numbers, strengthening the signal associated with grain size. Across six distinct metals, the same adjustment made the model favor higher strength when the grains got smaller and lower strength when they grew. Because the effect reversed with the physical situation, we could tell that the grain-size signal helped the model decide which answer to give.
The biggest opportunity I see is to turn these measurements into rewards for post-training. After its initial training, a model can learn from feedback. A reward is simply a score that reinforces a particular behavior. Imagine giving the model two versions of a problem, one in which the grains shrink and another in which they grow. We could give it credit for answering both correctly and additional credit when its internal response follows the physical law and helps produce both answers. Then we could change the metal, the numbers, or the wording and test whether the same reasoning holds. That makes the ability to carry a physical principle into a new situation something we can train for directly.
Coming from a multi-agent control background, I feel there is a lot in classical control (distributed consensus, game-theory based coordination, distributed methods for Dec-POMDP...) that can help guide multi-AI coordination
Multi-agent systems often behave like a mob, not like a coordinated team. We are releasing Multi Agent Coordination Bench: a benchmark requiring agent coordination, collaboration and communication. Right now, agents benefit from parallel exploration, but their coordination and communication is very poor. A team of eight agents does better than a single agent but often due to random exploration, while their coordination has significant room for improvement.
Our tasks include agent teams trying to execute a mission to Mars, fight zombies, grow a company from seed, and more. We categorize our tasks by two properties of the environment: if the agents share the same information and actions (Shared vs Distributed), and if they pursue the same goal or not (Collaborative/Strategic). This taxonomy will be useful for future research, to think about different multi-agentic problems across distributed and strategic settings.
Guannan Qu retweeted
Over the past week, I’ve been overwhelmed by the love and support of friends, industry colleagues, and people I've never met. Thank you to everyone who has shown kindness to my family and me. It means the world to me.
I came to the United States 53 years ago, and becoming an American citizen is one of the greatest gifts and proudest moments of my life. NVIDIA has been my life’s work. Building it into the company it is today alongside my colleagues has been a privilege and an honor. Nowhere but in America could a company like NVIDIA be possible. I love our country, and I love NVIDIA.
I did not live up to the standards expected of everyone at NVIDIA, and I have made serious mistakes in my personal life. I created problems for the Board that I deeply regret. I own my mistakes completely, and I am committed to cooperating with the Board’s independent investigation.
Since my departure, I’ve seen an outpouring of calls from NVIDIA’s employees, investors, business partners, and customers for me to return as CEO. Many continue to reach out to me directly as the company navigates complex technology, market, competitive, and geopolitical dynamics. NVIDIA is a once-in-a-generation company. If the Board of Directors decides that my return is in the best interest of our shareholders, our employees, the company, and America, I would be honored to answer these calls and return as CEO.
Thank you again to everyone who has shared their good wishes. I am profoundly humbled by your support. To my family: I love you, and I will spend the rest of my life being worthy of yours.
Jensen
very cool bench from colleagues at CMU (and others!)
Introducing Singularity-Bench 🕳️
🔄 We measure whether agents can autonomously conduct the full agent development lifecycle: six long-horizon, end-to-end tasks spanning pre-, mid-, and post-training, inference engines, cluster management, and harness development on production GB300 clusters!
We emphasize rigorous verification to guarantee correctness and execution integrity 👮♂️
🙌 Big shoutout to all authors contributing from academia and industry: @IronSteveZhou, @wonderingkrish, @chenzhuoming911, Zerui Wang, @KyriectionZhang, Wenshuang Li, Xinpeng Wei, Zhichen Zeng, @Jingyu227, @JueWANG26088228, @Wu_Sen, Ziheng Jiang, @zhendongucb, @hancai_hm, @Hades317, @ying11231, @ce_zhang, @uiuc_aisecure, and @BeidiChen!
How far are current agents from building their successors end to end? 🤔 We're here to find out!
🧵1/n
🌐 Website: infini-ai-lab.github.io/sing…
1
88
Proud of Baris Askin, @m_ustaomeroglu and @Annuppammm on our #NeurIPS2026 paper. A student model trained only on math, with a teacher that just grades each step, starts giving harmful advice on unrelated topics.
Our 5-min video explains it 👇
What we find:
- Misalignment follows the task more than the topic
- It transfers even under on-policy distillation
- The teacher decides where it goes; the data decides how much
Paper: arxiv.org/abs/2605.12798
With Gauri Joshi and Carlee Joe-Wong at CMU
1
70
Just back from #CoRL2026! At the workshop on GPU-accelerated sim to real-robot learning, I gave a talk on “The Power of Test-Time Sampling: From Simulators to World Models.”
With a simulator, the evaluator is exact, and sampling shines. We show sampling-based optimizers (CEM, MPPI, ...) are gradient descent on a smoothed objective, like score ascent in diffusion models. This gives global convergence guarantees.
arxiv.org/abs/2605.16520
1
1
119
Guannan Qu retweeted
LLMs Can In-context Learn Multiple Tasks in Superposition
We explore a bizarre LLM superpower that allows them to solve multiple ICL tasks in parallel.
This is related to the view of them as simulators in superposition [cref:@repligate]
arxiv.org/pdf/2410.05603
1/n
Guannan Qu retweeted
I guess this announcement is a bit (or super) late. I am on the 2023-2024 faculty job market for CS and data science!! Here is my humble homepage: zhundeng.org/home
Guannan Qu retweeted
Our next talk:
10/21: Tamer Başar (UIUC)
“Policy Optimization for Linear Optimal Control with Guarantees of Robustness”
Please visit sites.google.com/view/contro……… for more information (subscribe to the Google group to receive Zoom link and future announcements).
Guannan Qu retweeted
Upcoming speakers:
Babak Hassibi (Caltech)
Tamer Başar (UIUC)
Jeff Shamma (KAUST)
Sasha Rakhlin (MIT)
Melanie Zeilinger (ETH)
Claire Tomlin (UC Berkeley)
Jean-Jacques Slotine (MIT)
Na Li (Harvard)
Marco Pavone (Stanford)
Guannan Qu retweeted
Our first talk:
09/30: Elad Hazan (Princeton) @HazanPrinceton
"The Non-Stochastic Control Problem"
Please visit sites.google.com/view/contro… for more information (subscribe to the Google group to receive Zoom link and future announcements).














