Aleksandra Faust (@AleksandraFaust) header
Aleksandra Faust (@AleksandraFaust) profile picture

Foundation Models & AI Agents for the Real World | AI Innovation from Inception to Societal Impact | @genesismolai • @GoogleDeepMind • @Waymo • @SandiaLabs

Palo Alto, CA
Joined April 2009
Photo by Aleksandra Faust (@AleksandraFaust)Photo by Aleksandra Faust (@AleksandraFaust)Photo by Aleksandra Faust (@AleksandraFaust)Photo by Aleksandra Faust (@AleksandraFaust)Photo by Aleksandra Faust (@AleksandraFaust)Photo by Aleksandra Faust (@AleksandraFaust)Photo by Aleksandra Faust (@AleksandraFaust)Photo by Aleksandra Faust (@AleksandraFaust)Photo by Aleksandra Faust (@AleksandraFaust)Photo by Aleksandra Faust (@AleksandraFaust)
I’m humbled to share that I’ve been elevated to IEEE Fellow, Class of 2026! Citation: "For technical contributions to learning-based scalable autonomy and foundation models." Research is a team sport. This belongs to the incredible mentors & collaborators who shaped my path. ieee-org.widen.net/s/qqrqr2n…
5
32
8,433
Aleksandra Faust retweeted
Two Delphi panels (≈30 experts) defined consensus taxonomies of clinical safety risks and relational competencies. When stress-tested on 256 taxonomy-derived adversarial scenarios (concealed crises, pressure for unsafe prescriptions), training nearly eliminated contraindications and more than halved acute under-triage. Far fewer failures, yet still too many for autonomous care today.
1
1
4
228
Access to high-quality psychiatric care remains out of reach for far too many people. Could AI play a role in closing that gap? In new research, our team trained AI agents with reinforcement learning (RL) with simulated patients to conduct psychiatric assessments. Psychiatric evaluation isn't simple Q&A. It requires sustained interaction: eliciting information, prompting reflection, forming and refining clinical hypotheses, recognizing risks, and adapting the conversation as new information emerges. Four key takeaways: 1. Learning through dynamic interaction. Rather than just learning static answers, the agent used RL in simulated encounters to practice iterative information gathering, clinical reasoning, and patient-centered communication as the dialogue evolved. 2. Sim-to-human transfer. Crucially, improvements transferred beyond simulation. In a randomized, double-blind study involving psychiatrists role-playing patients, with independent psychiatrists evaluating the conversations, our RL-trained agent outperformed its base AI model in differential diagnosis appropriateness (68.3% vs. 33.3%) and patient-centered communication (82.0% vs. 65.2%). 3. Safety across trajectories. In healthcare, failing to probe a risk can be as dangerous as giving incorrect advice. In adversarial simulations, acute risk under-triage dropped from 30.5% to 12.9%. Safety isn't just about individual responses, but what happens throughout an interaction. 4. The simulation evaluation gap. LLM-based evaluations captured the overall direction of improvement but overestimated the magnitude of gains compared with human evaluations. Understanding where simulation predicts human-evaluated performance—and where it doesn't—is an important research challenge. The bigger picture: This approach points toward a scalable way to train agents for high-stakes, interactive domains without relying on trial and error with real people. What an agent learns depends not just on what it knows, but on what it asks, how it asks, and how people respond. An important caveat: These are early research findings, not real-world clinical trials. The human evaluations involved psychiatrists portraying patients, not actual patients. This is evidence of simulation-to-human interaction transfer, not readiness for independent clinical deployment. Still, it's an exciting step toward AI that learns to interact more effectively and responsibly—and a future where high-quality care is more accessible. Congratulations to the team. Paper: arxiv.org/abs/2610.40476
We trained Gemini via RL to conduct psychiatric intake with simulated patients. In a double-blind study with 32 psychiatrists playing the role of patients, we show that multi-turn RL in simulation improves Gemini’s performance on structured text-based psychiatric intake. 🧵
3
3
16
1,113
Aleksandra Faust retweeted
🎙️ Yet another position in Google DeepMind Tokyo 🇯🇵 私のチームでは Gemini Live 等の核となる音声対話技術、そしてAPAC拠点を活かした多言語・多文化LLMの研究を推進する Research Scientist/Engineer を募集中です。 Apply here: google.com/about/careers/app…
14
1
85
32,016
I'm hiring! 🚀 Our Frontier Health team at @GoogleDeepMind is looking for a Research Scientist to build foundational AI & physiological world models for human biology. Help us transform healthcare from reactive observation to proactive intervention. 🧬 🔗 Apply: goo.gle/3SPNis2
19
83
5
692
152,228
Thrilled to share our team's latest work! While LLMs ace static medical exams, real clinical reasoning requires navigating uncertainty over long, dynamic conversations. With ResidencyRL, we use multi-turn RL to train agents through simulated clinical practice. Massive congrats to @valentinlievin and the entire @GoogleDeepMind @GoogleResearch team on this huge milestone! 👇
Medical training begins in textbooks, but clinical competence is forged through practice. We introduce ResidencyRL: a multi-turn online RL method to train health agents in simulation, and demonstrate significant gains on top of frontier systems. arxiv.org/abs/2608.07418
1
2
31
2,912
Come join us and let's push the boundary of RL and agents to save lives.
Our posting for joining Google DeepMind as a Research Scientist was down for a few days but now it is back up! Apply here: google.com/about/careers/app… And fill out this form: docs.google.com/forms/d/e/1F…
1
1
11
4,478
We are entering a New Intelligence Age where technology is compressing decades of discovery into days. The ink we put on the paper won’t just write code—it will write the rules of human agency for the next century. Engineering alone is no longer enough. We need stewards. Delivering the convocation keynote for the @UNMengineering was an honor, but sharing the stage with my son as he delivered the student address made it uniquely unforgettable. 👇
1
22
1,325
My challenge to the Class of 2026: "Write a story that protects, democratizes, and elevates the human spirit. Make the work you do worthy of the power you wield." Full video, live transcript, and my deeper reflections on engineering stewardship in my latest Substack post: open.substack.com/pub/aleksa…
2
301
Happening now.
Excited to share my internship work with Genesis on biomolecular structure prediction at ICLR 🇧🇷! We simplified the structure predictor’s trunk, leading to 2-4x speedups! Find me at the Pairmixer poster: 🗓️ Saturday, 3:15–5:45 PM 📍 Pavilion 3, P3-#1022
1
9
1,513
Aleksandra Faust retweeted
We are honored to announce the Test of Time awards for #ICLR2026 🏆 This award recognizes papers published 10 years ago at ICLR 2016 that have had a lasting impact on the field: blog.iclr.cc/2026/04/22/anno…
2
50
17
331
102,676
Aleksandra Faust retweeted
Announcing the #ICLR2026 Outstanding Paper Awards 🏆 Congratulations to:
7
98
9
1,013
87,594
Aleksandra Faust retweeted
We’re looking forward to @iclr_conf in Rio de Janeiro. ➡️ Our Staff ML Research Scientist Pranav Murugan and former Genesis intern @zhang_ouyang, now a postdoc at MIT, will present findings from our Pairmixer biomolecular structure prediction model as a poster on Sat, Apr 25. More details: iclr.cc/virtual/2026/poster/…. We’ve also released the accompanying open-source model weights - check it out: github.com/genesistherapeuti…. ➡️ Our ML Research Scientist @RBoiarsky is co-organizing the Learning Meaningful Representations of Life (LMRL) Workshop on Mon, Apr 27: lmrl.org/ ➡️ Sign up for a coffee chat with a member of our team to learn more about our open roles in ML research: jobs.ashbyhq.com/genesis-mol… Hope to see you there!
4
2
5
1,809
Aleksandra Faust retweeted
ICLR 2026 is almost there! We have 6 exciting keynotes covering a range of areas from machine learning to robotics, neuroscience and AI for science: Maja Matarić, Max Welling, Percy Liang, Katie Bouman, Karen Adolph, Pablo Arbeláez See you all soon! #ICLR2026
2
28
2
291
39,683
Heading to @iclr_conf this week and I'll be there for the entire conference. Drop me a line or find me if you want to connect.
32
2,649
Aleksandra Faust retweeted
Calling robotics startups across Europe! Applications are now open for the @GoogleDeepMind Accelerator: Robotics, a 3-month program designed to help early-stage robotics companies scale from prototype to production. 🤖Powered by Google AI technologies, including Google Gemini Robotics Models. Apply today: goo.gle/robotics
We’re scaling up robotics in Europe. 🤖 Our Robotics Accelerator is tailored for startups and designed to bridge the gap between technology and business, powering the next generation of physical agents.
4
27
1
186
19,972
We are moving past the era of one-off digital assistants. For researchers, it’s been clear: we’re now building a self-improving learning engine. This isn't just about more compute; it’s about a fundamental shift that is (quietly) accelerating the path to AGI. Four specific pillars are enabling this transition: 1. Intent-based interaction: Natural language has broken the task barrier. Instead of rigid parameters, human intent now maps directly to machine action. 2. Code as actions: A universal layer that lets agents synthesize the interfaces they need on the fly, expanding their own capabilities to match any problem. 3. Generative simulation: Replacing manual engineering with high-fidelity sensory world models and "Social Digital Twins" to stress-test AI before it ever touches a real human. 4. Autonomous Self-Improvement: Models are beginning to teach themselves through experience—refining their own reasoning without waiting for human intervention. Together, they form a closed-loop system that breaks the traditional bottlenecks of AI development. Because tasks can expand via language and agents can "write" their own interfaces with code, they can explore increasingly diverse problems within generative worlds, and then refine their own knowledge without us labeling a single new data point. The result is an accelerated path to AGI—and on that path, I want to hold two ideas at once. First, the same shift that is accelerating AI capability is also reshaping our baseline for trust. As generative AI makes synthetic content indistinguishable from real, our global circle of trust—built over a century of mass media—begins to collapse. We are returning to something older: trust defined not by scale, but by depth. Direct observation. Shared history. Knowing who actually shows up and speaks in their own voice. Heck, we are starting to add misspellings in our own writing to indicate authencity! Second, the most important question for AI right now is not when AGI arrives. The most important question is whether we build AI as a force for genuine human augmentation, or as one more tool for synthetic noise (or worse). I’ve explored the technical and the societal implications in two new two deep-dives on Substack: aleksandrafaust.substack.com… I’d love to hear your thoughts.
3
1
1
25
2,872
1/4 I’m excited to share that I’m returning to @GoogleDeepMind as Research Director this week! Back to the lab to push the boundaries of agentic systems and AI for Health.
23
20
2
629
39,369
3/4 Huge thanks to @enfeinberg, @edunov and the team. The foundation is strong, and I’m excited to see where you take it next. In my advisory role, I'll support the leadership team to ensure continuity.
1
13
2,719
4/4 So why Google DeepMind now? My core research interests—self-improving models and agentic systems—are the key to solving humanity's hardest challenges. @GoogleDeepMind is the best place to push these technologies to truly serve those needs, particularly in #AI4Health.
2
26
1,770