We're graduate students, postdocs, faculty and scientists at the cutting edge of artificial intelligence research.
Berkeley, CA
Joined July 2017
- Tweets1.7K
- Following491
- Followers301K
- Likes758
Berkeley AI Research retweeted
I’m so excited to be giving an Early Career Keynote at #CoRL2026!
I’m especially grateful to the organizing team for creating a space for early-career researchers to share their work and perspectives!
Can’t wait to reconnect with everyone in the robot learning community!
🔥 #CoRL2026 is officially SOLD OUT! And I’m on my way to Austin! ✈️
What a journey this has been! Together with my amazing fellow Program Chairs, @Majumdar_Ani and @fox_dieter17849, we’ve spent the past year working to make this year’s CoRL something truly special.
📈 A record-breaking 2,094 submissions, with 687 accepted papers showcasing the incredible progress in robot learning!
🌟 An outstanding scientific program, featuring inspiring keynotes from @wolfram_burgard, @RussTedrake, @drfeifei and a special conversation with Fei-Fei Li and @LisaSu!
💡3 Early Career Keynotes to give the spotlight to the next generation of leaders in robot learning @RobobertoMM @negar_mehr @zhaohang0124!
🚀 New initiatives we’re particularly proud of:
- A Late-Breaking Results track to bring the freshest ideas and emerging directions into the conversation, including Agentic Robotics and Real2Sim2Real.
- A strengthened reviewing structure with Senior Area Chairs to help us manage the growing scale of our community while maintaining scientific rigor.
Behind all of this is an extraordinary team of organizers, reviewers, area chairs, volunteers, and researchers who have worked incredibly hard to make it happen. Thank you all! ❤️
🤝CoRL has grown tremendously, but what makes it special is still the same: a community that brings together bold ideas, rigorous science, and a shared passion for advancing robot learning.
I can’t wait to see everyone in Austin, hear the discussions, discover new ideas, and celebrate our community together!
We’ve put our hearts into this program. I really hope you enjoy it! ❤️
See you at @corl_conf! 🤖
Berkeley AI Research retweeted
I had my second baby ~6 months ago 😊 missing CoRL to send time with my family 🫶🏻 but here's a quick look at our lab's recent work, an extended version of last week's BARS talk!
My students are all at CoRL, please say hi! Enjoy👢🦿: youtube.com/watch?v=_nKhDrXV…
Berkeley AI Research retweeted
Don’t worry, robotics isn't solved yet bc they’re not socially intelligent at all.
Vongani took a stab at this by building on her previous human motion sequence model to allow humanoid to react to humans socially (with no mocap)! Vongani is on the market 🔥, find her at CoRL!
Berkeley AI Research retweeted
Introducing Social Humanoid: a framework for socially responsive human–humanoid interaction.
Deployed on a Unitree G1, our hierarchical system combines high-level social reasoning with low-level control, enabling a humanoid to interact with a human partner.
Berkeley AI Research retweeted
Happy to release a small report on some studies at Cursor on stabilizing the combination of RL + self-distillation in production-like agentic settings. w/ @srush_nlp and @ashvinair
arxiv.org/abs/2611.05266
This work/setup is a bit stale now, but could provide some insight on what pitfalls to lookout for!
Berkeley AI Research retweeted
Congratulations to the nine faculty receiving CDSS Machine Learning Research Awards! @UCBStatistics @Berkeley_EECS
@Cal_Engineer @UCBerkeley
🔗 Read more: bit.ly/4sHn3Sd
#MachineLearning #Statistics #ComputerScience #ResearchAwards #PeopleOfCDSS
ALT Grid of CDSS Machine Learning Research Award recipients Serina Chang, Avi Feller, Amanda Coston, Ryan Giordano, Jason Lee, Sewon Min, Narges Norouzi, Joseph Gonzalez, and Nikita Zhirotovskiy. Text reads: Nine CDSS faculty receive Machine Learning Research Awards.
Berkeley AI Research retweeted
A really fun project led by @KaiTyrusNelson, @tobykreiman: use simulated experiments to create a prior with state information, then align it to real-world experiments (with partial observation). Oddly enough, the theoretical foundations have a lot in common with inverse RL!
1/ A grand challenge in science: the simulation-to-experiment gap
Simulators model complex systems but introduce model errors; experiments give real-world data but are costly and noisy
Can we leverage both? Yes! We introduce Adversarial Distribution Alignment (ADA), at #NeurIPS2026
ADA aligns generative models—initially trained on simulations (e.g., molecular dynamics with classical force fields)—to full real-world distributions (e.g., protein structures) using only partial, noisy observations (like cryo-EM images)
ADA goes beyond just matching the expectation of observables: it can match full distributions, even in the presence of multiple correlated, high-dimensional observations
Check out our website: kaityrusnelson.com/ada/
And our paper: arxiv.org/abs/2510.02259
In collaboration with @KaiTyrusNelson, @tobykreiman, and @svlevine
Berkeley AI Research retweeted
AI is changing how we build software.
The next frontier is matter.
Meet CliqueFlowmer: upload a material, optimize its representation, generate a new candidate.
Our demo is live. Our paper is coming to #NeurIPS2026 in #Sydney.
Watch it happen. 🧵
Berkeley AI Research retweeted
Our JupyterHealth team won a $10mil @LaudeInstitute Moonshot Prize! We will build an open-source health data infrastructure to bring AI-powered healthcare into people's every day lives (aka "Edge Medicine") 🚀
I'll be hiring 1-2 PhD students this cycle, so please apply 🤩
Berkeley AI Research retweeted
Before you stack complex corrections in async RL, check out these findings: score centering alone removes drift and often matches on-policy performance!
🤔 Is score centering (SC) all we need for stable async RL?
People often stack truncated importance sampling (TIS) or KL penalties on top of SC. In a fun side project, we tested how much they help.
We found that SC alone often matched on-policy. But when it fell short, adding more corrections doesn't necessarily help. 🧵
Berkeley AI Research retweeted
🤔 Is score centering (SC) all we need for stable async RL?
People often stack truncated importance sampling (TIS) or KL penalties on top of SC. In a fun side project, we tested how much they help.
We found that SC alone often matched on-policy. But when it fell short, adding more corrections doesn't necessarily help. 🧵
Berkeley AI Research retweeted
What if you evolved the harness alongside the solution in an agentic loop? Big reductions in the cost and time to generate state-of-the-art solutions. Check out Exo, our recent work, for more details.
Agent harnesses are usually hand-tuned, or optimized offline against benchmarks. What if an agent could improve its own harness *while doing the work*?
Introducing Exo, a harness architecture for runtime self-improvement 🧵
(paper: arxiv.org/abs/2611.00463)
ALT Diagram: during a service incident, the agent starts with a policy that feeds full log history to the model and uses basic log/service tools. It self-edits mid-task, revising context assembly to summaries plus retrieved details and building a batched diagnostic tool, then continues the same incident with the revised harness.
Berkeley AI Research retweeted
CoT training often relies on task-specific supervision of intermediate steps. We introduce Abstract Token Curriculum (ATC), a curriculum-based method for training continuous CoT that reduces the need for intermediate supervision. Work with @khashagatmiry,@parsamirta (1/8)
Berkeley AI Research retweeted
1/ A grand challenge in science: the simulation-to-experiment gap
Simulators model complex systems but introduce model errors; experiments give real-world data but are costly and noisy
Can we leverage both? Yes! We introduce Adversarial Distribution Alignment (ADA), at #NeurIPS2026
ADA aligns generative models—initially trained on simulations (e.g., molecular dynamics with classical force fields)—to full real-world distributions (e.g., protein structures) using only partial, noisy observations (like cryo-EM images)
ADA goes beyond just matching the expectation of observables: it can match full distributions, even in the presence of multiple correlated, high-dimensional observations
Check out our website: kaityrusnelson.com/ada/
And our paper: arxiv.org/abs/2510.02259
In collaboration with @KaiTyrusNelson, @tobykreiman, and @svlevine
Berkeley AI Research retweeted
Agent harnesses are usually hand-tuned, or optimized offline against benchmarks. What if an agent could improve its own harness *while doing the work*?
Introducing Exo, a harness architecture for runtime self-improvement 🧵
(paper: arxiv.org/abs/2611.00463)
ALT Diagram: during a service incident, the agent starts with a policy that feeds full log history to the model and uses basic log/service tools. It self-edits mid-task, revising context assembly to summaries plus retrieved details and building a batched diagnostic tool, then continues the same incident with the revised harness.
Berkeley AI Research retweeted
A great example of using SkyDiscover to automatically build (synthesize) an inference engine from scratch for a specific model and a specific accelerator!
What if you could synthesize an LLM inference engine customized for your model and Mac, just in time?
Meet JITEngine: custom engines built with SkyDiscover for Apple Silicon, up to 2.7× faster than the fastest alternatives we tested.
jitengine.ai 🧵 (1/n)
Berkeley AI Research retweeted
What goes wrong when LLMs recursively self-summarize in long-horizon info gathering tasks? ⛓️💥 And how can this be fixed if we don't have expert demonstrations? Hint: RL with outcome rewards is not enough! See our paper now published @TmlrOrg ➡️ openreview.net/pdf?id=3j6wBs…
Berkeley AI Research retweeted
What if you could synthesize an LLM inference engine customized for your model and Mac, just in time?
Meet JITEngine: custom engines built with SkyDiscover for Apple Silicon, up to 2.7× faster than the fastest alternatives we tested.
jitengine.ai 🧵 (1/n)
Berkeley AI Research retweeted
When running RL with prior data, a behavior constraint is important to trade off exploration vs staying close to known good parts of the data. But the choice of this constraint is delicate -- we don't want to just match all the prior data, but we don't want to wander too far. In our new paper, we propose a new constraint that combines support constraints with continuous penalties to enable excellent offline to online RL performance.
While working on RLT (pi.website/research/rlt) @physical_int, it became clear to me that online RL needs the right constraints to guide exploration without hindering improvement.
Excited to introduce 🌟GSBR🌟, a new behavior regularizer that allows RL actor to both (1) drop bad modes and (2) explore optimal actions outside the base support.
(1/n) A thread below 🧵 👇
Berkeley AI Research retweeted
While working on RLT (pi.website/research/rlt) @physical_int, it became clear to me that online RL needs the right constraints to guide exploration without hindering improvement.
Excited to introduce 🌟GSBR🌟, a new behavior regularizer that allows RL actor to both (1) drop bad modes and (2) explore optimal actions outside the base support.
(1/n) A thread below 🧵 👇