📣📣📣 Join us for my group’s talks at #INFORMS2026 annual meeting. Two super star PhD students are on the job market and will present in the Job Market Showcase.

📚 What is 𝐚𝐧𝐚𝐥𝐲𝐭𝐢𝐜𝐬 𝐬𝐜𝐢𝐞𝐧𝐜𝐞, and can it lead to better solutions for society's hardest problems? I'm excited to join @INFORMS for a FREE virtual Book Talk on my book, "𝐈𝐧𝐬𝐢𝐠𝐡𝐭-𝐃𝐫𝐢𝐯𝐞𝐧 𝐏𝐫𝐨𝐛𝐥𝐞𝐦 𝐒𝐨𝐥𝐯𝐢𝐧𝐠: 𝐀𝐧𝐚𝐥𝐲𝐭𝐢𝐜𝐬 𝐒𝐜𝐢𝐞𝐧𝐜𝐞 𝐭𝐨 𝐈𝐦𝐩𝐫𝐨𝐯𝐞 𝐭𝐡𝐞 𝐖𝐨𝐫𝐥𝐝." 🎉 Data is everywhere, but insight is rare. Too often, AI and analytics produce impressive models or solutions that never change a decision. In this talk, I'll share how an 𝐢𝐧𝐬𝐢𝐠𝐡𝐭-𝐝𝐫𝐢𝐯𝐞𝐧 𝐚𝐩𝐩𝐫𝐨𝐚𝐜𝐡 can bridge that gap and help us tackle today's challenges and tomorrow's opportunities, from healthcare to public policy and beyond. 🌍 💡 What we'll discuss: 🔹 Why insight, not just data, drives real impact 🔹 How AI and analytics can address pressing societal problems 🔹 Where AI and analytics meet real-world decision-making 🔹 Your questions, in an open, live conversation 🗓️ Wednesday, October 7, 2026 ⏰ 12:00–1:00 p.m. ET 💻 Virtual (Zoom) 🎟️ Free and open to all Whether you're a researcher, student, practitioner, or policymaker, or just curious about using analytics to make the world better, I'd love to see you there. 🙌 👉 Register here: connect.informs.org/discussi… ♻️ Please share with colleagues and students who might find it valuable! #INFORMS #Analytics #OperationsResearch #DataScience #DecisionMaking #PublicPolicy #ProblemSolving #BookTalk #AI

🎉 After 11 incredible years @Harvard I'm starting a new chapter. I've sat with this post for a few months, wanting to get it right—because it marks the end of something wonderful and the start of something I'm 𝐠𝐞𝐧𝐮𝐢𝐧𝐞𝐥𝐲 𝐞𝐱𝐜𝐢𝐭𝐞𝐝 about. After 11 wonderful years at @Harvard I've joined @SmithSchool and the Artificial Intelligence Interdisciplinary Institute at Maryland (AIM) at @UofMaryland. These two appointments will let me pursue the intersections that matter most to me and expand the reach and impact of the work I care about. To my incredible colleagues, students, senior fellows, and postdocs at my lab—the @PIASLab @HarvardBizGov—thank you! I learned as much from you as I hope you learned from me, and these 11 years shaped me in ways I'm still discovering. 🙏 I won't pretend this was a simple decision. The political environment at @Harvard over the past few years affected many scholars, and it is never easy to get over unjust decisions. But I'm grateful to have found a new intellectual home, and I'm looking forward to learning from remarkable colleagues at Smith, AIM, and across the University of Maryland campus. Onward to what's next. 🚀

⏱️ 11 𝐦𝐢𝐧𝐮𝐭𝐞𝐬 (4.2%) reduction---> Real-world impact of analytics science with zero capacity investment. That's how much shorter the average ER visit became at @MayoClinic Clinic— with no new beds, no new hires, no new software, and no new IT integration. Just a better decision about where each patient should be seen, suggested by analytics science and proved via a 𝐩𝐫𝐨𝐬𝐩𝐞𝐜𝐭𝐢𝐯𝐞 𝐛𝐞𝐟𝐨𝐫𝐞-𝐚𝐧𝐝-𝐚𝐟𝐭𝐞𝐫 𝐟𝐢𝐞𝐥𝐝 𝐭𝐫𝐢𝐚𝐥. 🎉 I'm thrilled to share that our study, "Vertical Patient Streaming in Emergency Departments," is now published in 𝐌𝐚𝐧𝐚𝐠𝐞𝐦𝐞𝐧𝐭 𝐒𝐜𝐢𝐞𝐧𝐜𝐞. Here's why I think it matters. 👇 🏥 Most EDs already have a "vertical" area — a seated treatment space for patients who don't need to occupy a bed. But the decision about who goes there is usually made ad hoc, in the middle of a busy shift, with no consistent logic behind it. That variability quietly costs hospitals enormous capacity. So we asked a simple question: can we make that one decision better? 🤖 Our team analyzed nearly 50,000 ED visits to build a machine learning model that predicts — using only what's already captured at triage — whether a patient will ultimately need a bed. We fed those risk scores into a stochastic patient-flow model to derive the optimal routing policy. Then we did the hardest part: we compressed all of it into a decision tree a clinician can actually use at 2 a.m. 📊 Then we tested it in the real world — a 13-week prospective before-and-after field trial with 11,015 patients in Mayo Clinic Arizona's new emergency department: ✅ 11-minute (4.2%) reduction in total ED length of stay ✅ 8-minute (4.5%) reduction in time from arrival to disposition ✅ No increase in 72-hour return visits — safety fully maintained Scaled to a medium-sized ED seeing ~40,000 patients a year, that's roughly 6,800 bed-hours recovered annually — capacity for about 2,000 additional patients and an estimated $3M in additional reimbursement. Without a single new bed. 🛏️❌ 🙏 None of this happens alone. My deepest thanks to my brilliant co-authors from the @PIASLab and @MayoClinic: Arshya Feizi, Agni Orfanoudaki, Wayne Martini, and @nhodgsonem— to @INFORMS for highlighting our work and to the physicians, nurses, and triage staff who didn't just let us study their ED, but helped us redesign it. A model doesn't save a minute of anyone's time until a clinician trusts it enough to use it. 📄 Read the full paper in Management Science: pubsonline.informs.org/doi/1… #Analytics #MachineLearning #HealthcareAnalytics #OperationsResearch #EmergencyMedicine #PatientFlow #HealthcareInnovation #INFORMS #ManagementScience #PIASLab

Thank you @ICS_Informs for highlighting this paper (published by Management Science) as part of the @INFORMS Harvey J. Greenberg Research Award research highlights. Grateful to be included among the top three papers that were selected for the award. Full paper: pubsonline.informs.org/doi/a… #ReinforcementLearning #AI #CausalInference #CausalAI #MedicalAI

Can an LLM learn to make good sequential decisions? That's the question behind my latest #PIAS #blog post, "𝐏𝐫𝐞𝐩𝐚𝐫𝐢𝐧𝐠 𝐀𝐈 𝐟𝐨𝐫 𝐒𝐞𝐪𝐮𝐞𝐧𝐭𝐢𝐚𝐥 𝐃𝐞𝐜𝐢𝐬𝐢𝐨𝐧-𝐌𝐚𝐤𝐢𝐧𝐠 — 𝐚𝐧𝐝 𝐖𝐡𝐲 𝐈𝐭 𝐌𝐚𝐭𝐭𝐞𝐫𝐬." Most of the excitement around LLMs has focused on one-shot tasks: classification, translation, a clever answer to a single prompt. But the decisions that matter most in the real world don't work that way. In our recent work, we asked a simple question: can we equip a pretrained LLM to make good sequential decisions from offline data — and can we understand why it works? A few takeaways: 🔹 Instead of building a specialized model from scratch, we fine-tune an open-source pretrained LLM so it can read a handful of demonstrations and act in a new, unseen task — no further updates needed at test time. 🔹 We don't stop at intuition. For linear MDPs, we look inside the attention layer and show it behaves like an implicit estimator of the optimal value function, with a clean theoretical bound on how far the learned policy can fall short of optimal. 🔹 We tackle the hard, realistic cases — not just fully observable settings, but partial observability and genuine model ambiguity, which are the norm in medicine, not edge cases. 🔹 The gains are largest exactly where they're hardest to get. On a sparse-reward task with hidden goals, the approach reaches ~95% of oracle performance on held-out goals. If you care about where AI is heading — beyond chatbots and toward robust, data-efficient decision support in high-stakes domains like healthcare — I think you'll find this one worth a read. It sits right at the intersection of LLMs, reinforcement learning, and real clinical problems. 📖 Read the full post here: soroushsaghafian.com/prepari… I'd love to hear your thoughts — especially from those working on decision-making under uncertainty. #AI #MachineLearning #LLMs #Healthcare #ReinforcementLearning #DecisionMaking #ArtificialIntelligence

🚨 New research: Can LLMs be used for 𝐬𝐞𝐪𝐮𝐞𝐧𝐭𝐢𝐚𝐥 𝐝𝐞𝐜𝐢𝐬𝐢𝐨𝐧-𝐦𝐚𝐤𝐢𝐧𝐠 in complex, ambiguous environments? Our new paper — Large Language Models for Sequential Decision-Making: Improving In-Context Learning via Supervised Fine-Tuning — shows the answer is yes, and the gains are substantial. Most AI decision-making research assumes the world is fully observable and unambiguous. Real-world problems — especially in domains such as healthcare — are neither. Our framework fine-tunes a pretrained LLM (Llama-2-7B) on offline, oracle-labeled trajectories so it can tackle MDPs, POMDPs, and Ambiguous POMDPS (APOMDPs): settings with partial observability and model ambiguity. 💡 Key findings: → SFT slashes the optimality gap from 43% → 15% vs. random baselines in long-horizon MDPs → In partially observed and ambiguous environments, fine-tuned LLMs outperform ICL-only baselines and DPT by up to 15 percentage points → On the Darkroom navigation task, our model achieves ~95% of oracle performance → Robust to out-of-distribution conditions — strong generalization without retraining 📐 On the theory side, we interpret the fine-tuned attention mechanism as implicitly estimating optimal Q-functions, and derive an end-to-end suboptimality bound that cleanly separates in-context estimation error from training-length bias. This matters deeply for use in areas such as healthcare, where experimentation is costly or unethical and offline observational data are abundant. Our goal: give clinicians AI systems that reason under genuine uncertainty — not just clean textbook settings. Huge congratulations to my current and former lab members Minmin Zhang and @aghaei_sina on this work! 📄 Read the paper: arxiv.org/abs/2605.09009 #ArtificialIntelligence #MachineLearning #LLM #ReinforcementLearning #Healthcare #DecisionMaking #Research #Sequential

How can we make #AI impactful in #healthcare? Looking forward to this event hosted by @UICmedicine where I will be sharing some thoughts and answers.

“Let’s talk about data”📊 #Analytics #Sience (including #AI as one of its many branches) can help you improve the world and solve real-world problems. If you are interested, check out this short video and the book (Amazon: a.co/d/01AGVvdZ). 📺📺: youtube.com/shorts/MU9MS-WSO…