We're graduate students, postdocs, faculty and scientists at the cutting edge of artificial intelligence research.
Berkeley, CA
- Tweets 1,680
- Following 486
- Followers 298,242
- Likes 729
Berkeley AI Research retweeted
โHow can humanoids learn by watching humans?
Naive retargeting can lose crucial contacts: like bracing a hand on a table while bending to pick an object off the floor.
๐ Introducing VICAR: preserve key contacts from human videos, practice in simulation, deploy on hardware
1/n
8
23
86
15,730
Berkeley AI Research retweeted
New surprising result:
We trained AI recursively on its own safe generations, while giving it a benign cue. Models became increasingly misaligned over time, giving advice on how to do harm and build biological weapons.
We call this Progressive Misalignment. ๐งต๐
8
35
146
8,060
Berkeley AI Research retweeted
Last year I trained the advantage-conditioned VLA for ฯ*0.6 during my internship @physical_int. Now, I have a much improved recipe.
Introducing โจAdvantage Guidanceโจ, a performant RL method with stable advantage-conditioned supervised training and contrastive guidance sampling
5
35
334
25,589
Berkeley AI Research retweeted
Seeing is an act in itself. We look at what matters, and so should robots!
EyeRobot 2.0 does this with gaze, enabling precise manipulation without wrist cameras. Key for tool use, where the tool can block a wrist camera!
Co-led by @justkerrding who'll be on the job market! Hire him :)
Almost all bimanual robots include wrist cameras. They add bulk, cost, and complexity, but are critical for performance. Are they really needed?
Excited to finally share EyeRobot 2.0, where we use active gaze from an ego stereo camera instead ๐งต (with @kushtimusprime)
2
3
55
7,413
Berkeley AI Research retweeted
1/ We discovered something totally unexpected about the neural structure of consciousness that overturns decades of thinking on its biological basis and brings us significantly closer to the recipe for generating conscious awareness.
New preprint: biorxiv.org/content/10.64898โฆ (led by @HesseJanis, with @FrankLanfranchi, @_conor_mcgrory, @Yuelin_Shi, and @EngelTatiana)
384
1,681
11,277
2,140,406
Berkeley AI Research retweeted
Excited to be co-organizing our Trust4RSI-Agent Workshop at #AAAI2027 ! A Workshop on Trustworthy Evolution of AI Agents! ๐
๐ฅณWe have an amazing lineup of speakers and panelists: @Yoshua_Bengio , @MengdiWang10 , @du_yilun , @hugo_larochelle , @ZheqingZhu , and @__yuezhao__ !
๐ Submission deadline: November 20, 2026
๐ฌ Author notification: December 2, 2026
๐ February 22โ23, 2027 ยท Montrรฉal, Canada
๐ trust4rsi-agent.github.io/AAโฆ
Please share with colleagues working in this space and looking forward to your submissions!
Happy to be organizing this alongside @Haolun_Wu0203 @HuaWenyue31539 @BangL93 @lcharlin! #AAAI2027 #AIAgents
3
20
3,747
Berkeley AI Research retweeted
Can you sample a robot policy the way you sample an image? ๐ฒ
No RL run. No reward. No fine-tuning.
We train a diffusion model on the ๐๐ฒ๐ถ๐ด๐ต๐๐ of thousands of RL policy networks, then sample brand-new ones that walk ๐ถ, dance ๐บ, and spin-kick ๐ฆต on a real Unitree G1 ๐ค
The best part? The ๐๐ฒ๐ถ๐ด๐ต๐๐ have a secret ๐ Every network, no matter the motion, pays the most attention to the hips and waist โ embodiment comes first; motion comes second. ๐ฆพ
๐ arXiv: arxiv.org/abs/2610.20606
๐ Website: humanoid-weight-diffusion.giโฆ
#Robotics #HumanoidRobots #DiffusionModels #RobotLearning #EmbodiedAI
20
39
281
16,615
Berkeley AI Research retweeted
As agents perform AutoResearch they get fixated on their old ideas and do only incremental innovations. We introduce a simple context management trick that forces the agents to forget everything they have done except the best solution they have found so far.
This simple modification in context management allows the agents to get out of their idea basins and come up with more novel solutions.
Closely related to the work 'Fork and Flush' and 'Idea-basins' that @DimitrisPapail posted a few days ago.
Autoresearch with coding agents can now run for days on a single problem, but we've seen an increasing number of cases where they stop improving early. We address this issue with a simple idea: forgetting.
Introducing our new paper: "Scaling Discovery Requires Forgetting"
16
40
443
27,075
Berkeley AI Research retweeted
Gemini-4 Argon shows a large improvement in robotics understanding on RoboVista.
It reaches 81.9%, beats GPT-6 Astra and Claude Fable 5.1 in our evaluation with lowest eval cost. Compared with Gemini 3 Flash, it improves across every robotics domain. Example VQA๐
๐คHow well do today's VLMs actually understand real-world robotics? ๐
Excited to share RoboVista at #RSS2026 โ a systematic evaluation and benchmark for VLMs across diverse, real-world robot applications. Website, dataset and paper: berkeleyautomation.github.ioโฆ
Developed by researchers at @UCBerkeley, @GoogleDeepMind, and @Princeton. ๐งต๐
2
9
45
6,416
Berkeley AI Research retweeted
Slow evaluations are a major bottleneck in AI-driven systems optimization. DuoDiscover addresses this by having agents generate fast evaluation proxies (โyardsticksโ) alongside the code. The result: a 2.68ร speedup for TPU kernels and a 1.92ร speedup for vLLM serving!
๐ Excited to introduce ๐๐ฎ๐จ๐๐ข๐ฌ๐๐จ๐ฏ๐๐ซ!
This isnโt benchmark gaming. Itโs AI agents shipping real PRs and beating human-expert hardware baselines. ๐ฅ
๐๐ก๐ ๐๐ซ๐๐๐ค๐ญ๐ก๐ซ๐จ๐ฎ๐ ๐ก๐ฌ ๐
๐ ๏ธ Real optimizations merged into main branches (e.g., FreeToken #619)
๐ง First superhuman CPU prefetcher: +3.93% | +6.31% overall, outperforming AlphaEvolve
โก 2.68ร TPU kernel speedup | 1.92ร faster vLLM serving | ๐ #1 on NVIDIA SOL-ExecBench (B200)
๐ก ๐๐ก๐ ๐ฌ๐๐๐ซ๐๐ญ ๐ฌ๐๐ฎ๐๐? Agents co-discover fast evaluation proxies (โyardsticksโ) to slash slow eval bottlenecks and compound real performance gains at scale.
๐ค Incredible collaboration across @GoogleCloud, @BerkeleySky & @StanfordAILab!
๐ฅ Weโre pushing the frontier of AI-driven systems optimization. And weโre just getting started. ๐
2
6
84
7,399
Berkeley AI Research retweeted
So excited to release DuoDiscover! I worked on it over the summer for my internship @Google with some awesome collaborators. We got some amazing results on open source benchmarks and internal tasks!
Was super fun working closely with @shagarw21 @dilarafsoylu @MikeHeddes @RuiMeng_ @VioletNPeng and thanks to @chunliang_tw for being the most amazing manager ever!!
Such a cool collaboration between @googlecloud @berkeley_ai @StanfordAILab
AI is cracking math problems. What's holding it back in engineering?
The cost of evaluation. Agents can generate ideas in seconds, but testing them can take hours or days.
๐ก Meet DuoDiscover: where evaluation is part of what agents discover ๐๐งต
Some highlights:
๐ #1 on NVIDIA's SOL-ExecBench
๐ 2ร the experts on SWE-fficiency
๐ 1.92ร on vLLM; 1.44ร on FreeToken
๐ฅ 2.68ร over production TPU kernels
๐ฅ๏ธ Beats the best human and AI-designed CPU prefetchers
1
5
18
3,260
Berkeley AI Research retweeted
Excited and grateful to receive a 2026 AFOSR Young Investigator Program (YIP) award! ๐
Our project explores how machines can reason about human intentions, beliefs, and goals for better human-machine interactions ๐ค๐ง ๐ฉ.
Thank you AFOSR for supporting this work!
7
4
55
4,776
Berkeley AI Research retweeted
Cool results making auto-research / AI-based optimization better by having agents forget! Turns out that just running an agent for a long time can lock it into small steps in the same direction.
Autoresearch with coding agents can now run for days on a single problem, but we've seen an increasing number of cases where they stop improving early. We address this issue with a simple idea: forgetting.
Introducing our new paper: "Scaling Discovery Requires Forgetting"
1
15
111
17,260
Berkeley AI Research retweeted
Autoresearch with coding agents can now run for days on a single problem, but we've seen an increasing number of cases where they stop improving early. We address this issue with a simple idea: forgetting.
Introducing our new paper: "Scaling Discovery Requires Forgetting"
37
69
596
69,834
Berkeley AI Research retweeted
Thanks to Google.org for supporting our work with @oziadias using sparse autoencoders and agentic approaches to accelerate discovery in cardiology and oncology! A wonderful collaboration with @rajivmovva, @dmshanmugam, Alex Schubert, and others.
UC Berkeley researchers have been awarded $1.7M from Google.org to advance medical discovery using AI. @2plus2make5 @oziadias @Berkeley_EECS @UCJointCPH @googleorg @UCBerkeley
๐ Read more: bit.ly/4hUBn4H
#UCBerkeley #PeopleOfCDSS #AIforScience #Health
ALT Emma Pierson and Ziad Obermeyer. Text reads: UC Berkeley researchers awarded $1.7M from Google.org to advance medical discovery using AI.
7
6
105
7,726
Berkeley AI Research retweeted
Now autoregressive video models can remember. ๐ง
Introducing MosaiChunk: persistent visual memory for long-horizon autoregressive video generation๐งฉ.
Instead of remembering everything, it remembers the right things. โ๏ธ
1/6
5
59
290
30,905
Berkeley AI Research retweeted
We gave Claude and GPT a phone video of a robot wiping. In an hour, each built a full-dynamics digital twin.
Introducing RoboSimStudio ๐ค: an open and interactive toolkit for Agentic Robotics, enabling Fable, Astra and Argon AI agents to model sponges ๐งฝ, cables ๐งต and cloth ๐.
38
49
427
166,080
Berkeley AI Research retweeted
Hereโs a new โจoff-policyโจ addition to our series on RL for flow policies.
Idea is simple: hill-climb on โQ, filter for stability, flow-match the replay buffer. Sample efficient and works on humanoids, VLAs, and images!
Led by amazing @ChungMinKim, whoโs on the job market!
Excited to introduce QF3: Fast Flow RL with Filtered Q-Gradients!
Train humanoids from scratch, fine-tune VLAs, even fine-tune image models... all with one simple off-policy update โจ
2
5
69
6,779
Berkeley AI Research retweeted
Pushing TPUs to their limits. FLAT hits 90% of peak HBM bandwidth (1.48 TB/s) and delivers large training speedups for Mamba-2 and GLA over standard H100 baselines. Great deep dive into kernel-level hardware optimization!
Releasing FLAT: Flash Linear Attention on TPUs ๐
Pallas training kernels for Linear Attention, Mamba-2 and GLA. On one TPU v6e, forward + backward is up to 3x faster than FLA on an H100 for GLA, and 1.5x faster for Mamba-2.
๐งต
2
11
117
7,834
