First year PhD at @StanfordAILab advised by @chelseabfinn. prev harvard / rhodes
Joined May 2020
- Tweets288
- Following808
- Followers853
- Likes5.2K
Pinned Tweet
Introducing Real-time RL.
In the real world, time isn't free. The environment keeps "moving" even when you're computing your next action.
We show how RL agents can learn to adaptively think in real-time games.
1/🧵
Aneesh Muppidi retweeted
In principle, a hierarchical policy should facilitate both credit assignment and transferability.
In practice, HRL algorithms either discard information necessary for control, or retain unnecessary information, limiting experience aggregation for learning.
Introducing QSO...
Aneesh Muppidi retweeted
In language models, turning a generative model into a critic is simple: pass in the output and prompt for a score. Robot policies can't do this - they output physical actions & there's no built-in vocabulary to predict scores.
So when it comes to RL on robot policies, we typically train a separate critic from sparse rewards - blind to everything the policy learned from the actions in the pretraining data.
We think this has it backwards - can a critic learn from those actions the way the policy does? (1/n)
Aneesh Muppidi retweeted
Last year I trained the advantage-conditioned VLA for π*0.6 during my internship @physical_int. Now, I have a much improved recipe.
Introducing ✨Advantage Guidance✨, a performant RL method with stable advantage-conditioned supervised training and contrastive guidance sampling
Aneesh Muppidi retweeted
Excited to announce Next-Concept Modeling (NCM): our new paradigm for fast and parallel, coarse-to-fine text generation! 🚀
Compared to strong diffusion language models (DLMs) at similar parameter sizes, NCM showcases:
⚡ 31× faster generation on downstream summarization and translation tasks.
🏎️ 10–21× faster unconditional generation across TinyStories, LM1B, and OpenWebText, with comparable or better generative perplexity using 3–5× fewer training tokens.
🧩 Better reasoning performance on tasks like Sudoku.
Check out the paper and codebase 👇
Arxiv: arxiv.org/abs/2610.14162
AlphaXiv: alphaxiv.org/abs/2610.14162
Code: github.com/HazyResearch/ncm
Enjoyed working with @KumbongHermann and advised by @simran_s_arora @Azaliamirh and @HazyResearch.
🚀 Introducing Next-Concept Modeling (NCM): our new paradigm for fast, parallel, coarse-to-fine text generation.
Compared to strong diffusion language models (DLMs) at similar parameter size, NCM achieves:
⚡ Up to 31× faster generation on downstream tasks like summarization and machine translation with comparable or better quality.
🏎️ 10–21× faster unconditional generation across TinyStories, LM1B, and OpenWebText, with comparable or better generative perplexity and using 3–5× fewer training tokens.
🧩 Better performance on reasoning tasks like Sudoku, outperforming both AR and diffusion baselines with fewer sampling steps.
With @ishanskhare, @IddahMlauzi, @ekellbuch, @simran_s_arora, @Azaliamirh, and @HazyResearch.
More details in the thread 🧵👇
Aneesh Muppidi retweeted
Want to know how well agents can really control robots directly?
In RoboARCade, we systematically evaluate and diagnose frontier agents controlling robots directly using tens of thousands of dollars’ worth of agent trajectories. RoboARCade contains diverse, precise, and long-horizon tasks. Paper, code, assets and every trace are out today. roboarcade.org
Check out our new work! @larsankile led this amazing project, and I learned so much about how we can use deployment data to efficiently train robust policies. It’s a really practical guide, and anyone training policies should definitely give it a read.
I've been thinking a lot about making robot learning scalable. To me, that means learning on the job, and using the human help still needed efficiently.
In Mulligan ⛳️, we focus each data-collection round on the initial states where the robot fails. This happens in deployment, with the robot doing the task and a person stepping in when needed.
On 3 real tasks and 2,550 blind evals, it improves on uniform collection at the same budget.
Aneesh Muppidi retweeted
Today I'm excited to show off the HiveMind. I think it'll be the social media app that will connect the world; it'll allow you to understand anyone as well as you understand yourself.
It feels like the world is becoming more divided than ever; and it's because of social media. No one posts because you're scared to show your true self or don't have the time to document your life. Most platforms have devolved into short form dopamine machines.
The HiveMind is different. Though we share the same fundamentals as every social media app (stay caught up with your friends's lives and meet new people), we're a lot better at everything humans need to be social.
We make it easy to capture every aspect of your life through your camera roll. There's a notes tab (that feels like your favorite notes app) and allows you to talk into it (like your favorite dictation or meeting recording tool). And we give you an AI, your second brain, that talks to you, helps you, and documents your memories as you go.
You can then take those memories that are private to you, and choose to share it with different circles in your network. Your interactions with the Hive (the social feed) are also recorded as memories, which your AI second brain helps maintain.
Anything you see on the Hive, your AI can see as well. This allows your AI to help find you more cool people to talk to, or keep you updated on what your friends have been doing. And you can loop in your friends's AI into conversations, or message them directly to get caught up on their life.
I'd love it if you could give it a try! It's live on the app store, and if you'd be willing to hop on a call so I can see your experience, just comment and I'll send you a message!
I'm competing on a startup gameshow hosted by @stanleybystan #StanLaunchpad, and produced by the same people that make Mr. Beast videos. You'd be helping me survive elimination if you allowed me to interview you :)
Aneesh Muppidi retweeted
Introducing Real-Time EXPO-FT – Fast and Reliable RL for Real-Time VLA Policies!
Real-Time EXPO-FT unlocks π0.5 on challenging dynamic tasks, such as balancing a ball on a plate and striking a ball into the goal
(1/6)
Aneesh Muppidi retweeted
Excited to share Context-Sharded Block Parallelism (CSBP) and Turbo-dLLM, our new optimized distributed training library for Diffusion LLMs!
Our distributed parallelism strategy unlocks significant training efficiency for diffusion LLMs, with speedup gains growing with context length 🚀
On 8x H100 GPUs, Turbo-dLLM accelerates DFlash2 speculative-decoder training by 2.48x at 512K and 7.59x at 1M context length.
Simply pip install turbo-dllm⚡️ or check out scalingintelligence.stanford….
Honored to work with @TarunSures41845, @PranshuChatur11, @hangoo_kang, @pshroff_ , @KumbongHermann, and advisor @Azaliamirh 🌟!
Diffusion LLMs and speculative decoding promise much faster agents. Yet agents need long contexts, and training on them is painfully slow.
Introducing Context-Sharded Block Parallelism (CSBP), a new distributed parallelism strategy unlocking significant training efficiency for diffusion LLMs, with speedup gains growing with context length 🚀
⚡ 7.59× faster DFlash2 speculative decoding drafter training
⚡ 1.61× faster block diffusion fine-tuning
⚡ 1.33× faster autoregressive → block diffusion adaptation
With the same GPU hours, models trained with CSBP score higher on SWE-bench Verified and Terminal-Bench Lite 👑
Open-sourced in Turbo-dLLM, our new optimized distributed training library.
Advised by @Azaliamirh and with an amazing team: @PranshuChatur11 @hangoo_kang @pshroff_ @ishanskhare @KumbongHermann
Aneesh Muppidi retweeted
Excited to share SceneAgent 🤖, our agentic pipeline for scene simulation from 3D captures! Made by Luke Hollis and Tianxing Fan!
We use a combination of semantic features and predictive per-Gaussian physics to transform individual complex objects or entire scenes for robotics policy training and evaluation.
Our pipeline automates comprehensive details converting input images, video, lidar, or generated 3DGS scenes like from World Labs Marble.
It automates:
0. 3DGS processing from input data at 30k and 60k training steps, with automated per-image correction/calibration
1. Infer semantic features for each gaussian
2. Segment objects and infill background
3. Bake predictive physics materials for each object (rigidity, friction, density, etc).
4. Decompose objects into individual parts
5. Articulate movable pieces and joints
6. Generate similar 3d meshes with different geometries, textures, physics properties, and articulations
Connected to autoresearch tools, we believe that this will solve an important step in the time-intensive data problem for robotics training and policy evaluation.
Paper and code coming soon, more at computationalrobotics.seas.h…
Aneesh Muppidi retweeted
Simulating the consequences of an experiment before running it is a holy grail of AI for biology. How can we get there?
Our new paper, “World models for biomedicine,” which I led together with @njwfish, is out now in @CellCellPress’s special issue on AI in biology. We explore how biomedical world models could help scientists navigate the enormous space of possible biological interventions. In principle, they could:
🧪 Simulate how biological systems respond to interventions never previously tested
🎯 Search for interventions that drive cells, tissues, or patients toward desired states
🧭 Plan experimental campaigns over long time horizons
🤖 Power autonomous, closed-loop AI scientists
However, realizing this vision requires much more than calling a model a “world model.” What exactly defines a biomedical world model, and what capabilities distinguish it from related models? Where do today’s models already show promise, and where do they fall short? What new data and architectures are needed? What evaluations would demonstrate that a world model works?
We tackle these and other questions in this latest paper from @marinkazitnik's lab: cell.com/cell/fulltext/S0092…
More below 🧵👇🏽 1/8
Aneesh Muppidi retweeted
is Astra really this good at VLA?
I spent hours trying a zero-shot demo with the SO-101 arm and this is all I got
Aneesh Muppidi retweeted
Introducing OM-1, our first robot foundation model, zero-shot generalizing to any robot: table-top arms, industrial arms and humanoids.
- learned directly from human manipulation data
- no teleop/robot data
- close to human-level dexterity and efficiency
- multi-robot collab
Aneesh Muppidi retweeted
We built a small "primitive" of an artificial brain.
Think of primitives as Lego bricks for a different kind of intelligence, or more scientifically, something closer to the neural priors we're born with.
The really cool thing: the primitive wasn't trained. No learned policy. No reinforcement learning. No hand-coded obstacle avoidance (obviously). Just ~300 neurons reacting to a changing environment in real time through their own neural dynamics (that's all).
As we better understand some of the math and dynamics behind biological brains, we are now on a path to a different kind of robotic intelligence: running locally, without the need of millions of data points or traditional split between training and inference, and ultimately, able to react and learn continuously in world that changes.
We've only scratched the surface, but while our first papers go through peer review and our experiments continue, we'll be sharing more about primitives and two other core findings that gets us closer to continual learning. Follow if intrigued.
Note: We haven't solved continual learning yet, but now we're confident we can solve the basic dynamics behind it within the next year or so.
In the video (1x): ~300 no-pretrained neurons responding and adapting in real time using only their neural dynamics to control directly an e-puck body to avoid obstacles.
Aneesh Muppidi retweeted
We’ve all been amazed by how capable Astra is—whether at solving Millennium Prize–level mathematics, reconstructing scenes, or controlling robots.
A common pattern across these examples seems to be that, before Astra, many humans had already spent years attacking these problems—or closely related ones—creating a large body of knowledge, techniques, and examples. That accumulated human effort may be what eventually enables the tipping point where Astra becomes better, perhaps even substantially better, than individual humans at solving the problem.
By contrast, for problems that are not yet well defined, whose goals are ambiguous, or are so ill posed that very few people have seriously studied them, Astra still seems much less capable. There are research problems in my own group that I simply cannot imagine asking Astra to solve directly—not necessarily because the underlying mathematics is harder, but because we cannot yet formulate the problem clearly.
This creates an interesting dilemma. If you work on a hot and practically important problem, there is a good chance that Astra will soon be better at solving it than you are. If you work on something extremely niche, you may remain better than Astra—but the problem itself may not matter very much.
Perhaps the best research strategy, then, is what great research has always been: find an important problem that has not even been properly defined yet; discover the right way to formulate and frame it; and then use systems like Astra to help solve it.
The most valuable human contribution may increasingly shift from solving well-defined problems to discovering which problems should exist in the first place—and framing them in a way that, once solved, produces unexpectedly large impact.
Aneesh Muppidi retweeted
AI has solved Navier-Stokes. What would it take for AI to make similar advances in life sciences?
In our latest preprint, led by @AdaFang_, we argue that, while fields like mathematics and programming can cheaply evaluate many AI-generated candidates at scale, "closing the loop" of hypothesis ➡️ experiment ➡️ revised hypothesis remains a central bottleneck in AI-driven biomedical discovery.
We outline a vision for autonomous AI scientists that:
🧠 Reason over scientific knowledge, multimodal data, competing hypotheses, and uncertainty across multi-loop campaigns lasting days to months.
🔬 Learn under sparse and delayed feedback, combining soft verifiers (e.g., simulations, predictive models, and biological world models) with hard verifiers such as wet lab assays, robotic experiments, organoids, and clinical studies.
🎯 Allocate limited experimental budgets to tests with the greatest expected information gain.
We suggest that achieving this vision will require advances in long-horizon reasoning, process-based evaluation, uncertainty-aware exploration, self-driving laboratories, and human-on-the-loop oversight.
Read our paper here: preprints.org/manuscript/202…
More details below from @marinkazitnik. 👇🏽
Closed-loop AI scientists
AI can generate hypotheses much faster than we can experimentally verify them. @AdaFang_
In our latest piece on Closing the Loop in AI-Driven Biomedical Discovery we outline a path for closed-loop AI scientists, preprints.org/manuscript/202…
AI scientists can generate hypotheses, propose experiments, and analyze data, and several have produced findings that laboratories then confirmed.
What would it take to close that loop more generally?
