All posts by admin

The early years of faculty members’ careers are a formative and exciting time in which to establish a firm footing that helps determine the trajectory of researchers’ studies. This includes building a research team, which demands innovative ideas and direction, creative collaborators, and reliable resources. 

For a group of MIT faculty working with and on artificial intelligence, early engagement with the MIT-IBM Watson AI Lab through projects has played an important role helping to promote ambitious lines of inquiry and shaping prolific research groups.

Building momentum

“The MIT-IBM Watson AI Lab has been hugely important for my success, especially when I was starting out,” says Jacob Andreas — associate professor in the Department of Electrical Engineering and Computer Science (EECS), a member of the MIT Computer Science and Artificial Intelligence Laboratory (CSAIL), and a researcher with the MIT-IBM Watson AI Lab — who studies natural language processing (NLP). Shortly after joining MIT, Andreas jump-started his first major project through the MIT-IBM Watson AI Lab, working on language representation and structured data augmentation methods for low-resource languages. “It really was the thing that let me launch my lab and start recruiting students.” 

Andreas notes that this occurred during a “pivotal moment” when the field of NLP was undergoing significant shifts to understand language models — a task that required significantly more compute, which was available through the MIT-IBM Watson AI Lab. “I feel like the kind of the work that we did under that [first] project, and in collaboration with all of our people on the IBM side, was pretty helpful in figuring out just how to navigate that transition.” Further, the Andreas group was able to pursue multi-year projects on pre-training, reinforcement learning, and calibration for trustworthy responses, thanks to the computing resources and expertise within the MIT-IBM community.

For several other faculty members, timely participation with the MIT-IBM Watson AI Lab proved to be highly advantageous as well. “Having both intellectual support and also being able to leverage some of the computational resources that are within MIT-IBM, that’s been completely transformative and incredibly important for my research program,” says Yoon Kim — associate professor in EECS, CSAIL, and a researcher with the MIT-IBM Watson AI Lab — who has also seen his research field alter trajectory. Before joining MIT, Kim met his future collaborators during an MIT-IBM postdoctoral position, where he pursued neuro-symbolic model development; now, Kim’s team develops methods to improve large language model (LLM) capabilities and efficiency. 

One factor he points to that led to his group’s success is a seamless research process with intellectual partners. This has allowed his MIT-IBM team to apply for a project, experiment at scale, identify bottlenecks, validate techniques, and adapt as necessary to develop cutting-edge methods for potential inclusion in real-world applications. “This is an impetus for new ideas, and that’s, I think, what’s unique about this relationship,” says Kim.

Merging expertise

The nature of the MIT-IBM Watson AI Lab is that it not only brings together researchers in the AI realm to accelerate research, but also blends work across disciplines. Lab researcher and MIT associate professor in EECS and CSAIL Justin Solomon describes his research group as growing up with the lab, and the collaboration as being “crucial … from its beginning until now.” Solomon’s research team focuses on theoretically oriented, geometric problems as they pertain to computer graphics, vision, and machine learning. 

Solomon credits the MIT-IBM collaboration with expanding his skill set as well as applications of his group’s work — a sentiment that’s also shared by lab researchers Chuchu Fan, an associate professor of aeronautics and astronautics and a member of the Laboratory for Information and Decision Systems, and Faez Ahmed, associate professor of mechanical engineering. “They [IBM] are able to translate some of these really messy problems from engineering into the sort of mathematical assets that our team can work on, and close the loop,” says Solomon. This, for Solomon, includes fusing distinct AI models that were trained on different datasets for separate tasks. “I think these are all really exciting spaces,” he says.

“I think these early-career projects [with the MIT-IBM Watson AI Lab] largely shaped my own research agenda,” says Fan, whose research intersects robotics, control theory, and safety-critical systems. Like Kim, Solomon, and Andreas, Fan and Ahmed began projects through the collaboration the first year they were able to at MIT. Constraints and optimization govern the problems that Fan and Ahmed address, and so require deep domain knowledge outside of AI. 

Working with the MIT-IBM Watson AI Lab enabled Fan’s group to combine formal methods with natural language processing, which she says, allowed the team to go from developing autoregressive task and motion planning for robots to creating LLM-based agents for travel planning, decision-making, and verification. “That work was the first exploration of using an LLM to translate any free-form natural language into some specification that robot can understand, can execute. That’s something that I’m very proud of, and very difficult at the time,” says Fan. Further, through joint investigation, her team has been able to improve LLM reasoning­ — work that “would be impossible without the IBM support,” she says.   

Through the lab, Faez Ahmed’s collaboration facilitated the development of machine-learning methods to accelerate discovery and design within complex mechanical systems. Their Linkages work, for instance, employs “generative optimization” to solve engineering problems in a way that is both data-driven and has precision; more recently, they’re applying multi-modal data and LLMs to computer-aided design. Ahmed states that AI is frequently applied to problems that are already solvable, but could benefit from increased speed or efficiency; however, challenges — like mechanical linkages that were deemed “almost unsolvable” — are now within reach. “I do think that is definitely the hallmark [of our MIT-IBM team],” says Ahmed, praising the achievements of his MIT-IBM group, which is co-lead by Akash Srivastava and Dan Gutfreund of IBM.

What began as initial collaborations for each MIT faculty member has evolved into a lasting intellectual relationship, where both parties are “excited about the science,” and “student-driven,” Ahmed adds. Taken together, the experiences of Jacob Andreas, Yoon Kim, Justin Solomon, Chuchu Fan, and Faez Ahmed speak to the impact that a durable, hands-on, academia-industry relationship can have on establishing research groups and ambitious scientific exploration.

Characterized by weakened or damaged heart musculature, heart failure results in the gradual buildup of fluid in a patient’s lungs, legs, feet, and other parts of the body. The condition is chronic and incurable, often leading to arrhythmias or sudden cardiac arrest. For many centuries, bloodletting and leeches were the treatment of choice, famously practiced by barber surgeons in Europe, during a time when physicians rarely operated on patients. 

In the 21st century, the management of heart failure has become decidedly less medieval: Today, patients undergo a combination of healthy lifestyle changes, prescription of medications, and sometimes use pacemakers. Yet heart failure remains one of the leading causes of morbidity and mortality, placing a substantial burden on health-care systems across the globe. 

“About half of the people diagnosed with heart failure will die within five years of diagnosis,” says Teya Bergamaschi, an MIT PhD student in the lab of Nina T. and Robert H. Rubin Professor Collin Stultz and the co-first author of a new paper introducing a deep learning model for predicting heart failure. “Understanding how a patient will fare after hospitalization is really important in allocating finite resources.”

The paper, published in Lancet eClinical Medicine by a team of researchers at MIT, Mass General Brigham, and Harvard Medical School, shares results from developing and testing PULSE-HF, which stands loosely for “Predict changes in left ventricULar Systolic function from ECGs of patients who have Heart Failure.” The project was conducted in Stultz’s lab, which is affiliated with the MIT Abdul Latif Jameel Clinic for Machine Learning in Health. Developed and retrospectively tested across three different patient cohorts from Massachusetts General Hospital, Brigham and Women’s Hospital, and MIMIC-IV (a publicly available dataset), the deep learning model accurately predicts changes in the left ventricular ejection fraction (LVEF), which is the percentage of blood being pumped out of the left ventricle of the heart.

A healthy human heart pumps out about 50 to 70 percent of blood from the left ventricle with each beat — anything less is considered a sign of a potential problem. “The model takes an [electrocardiogram] and outputs a prediction of whether or not there will be an ejection fraction within the next year that falls below 40 percent,” says Tiffany Yau, an MIT PhD student in Stultz’s lab who is also co-first author of the PULSE-HF paper. “That is the most severe subgroup of heart failure.” 

If PULSE-HF predicts that a patient’s ejection fraction is likely to worsen within a year, the clinician can prioritize the patient for follow-up. Subsequently, lower-risk patients can reduce their number of hospital visits and the amount of time spent getting 10 electrodes adhered to their body for a 12-lead ECG. The model can also be deployed in low-resource clinical settings, including doctors offices in rural areas that don’t typically have a cardiac sonographer employed to run ultrasounds on a daily basis.

“The biggest thing that distinguishes [PULSE-HF] from other heart failure ECG methods is instead of detection, it does forecasting,” says Yau. The paper notes that to date, no other methods exist for predicting future LVEF decline among patients with heart failure.

During the testing and validation process, the researchers used a metric known as “area under the receiver operating characteristic curve” (AUROC) to measure PULSE-HF’s performance. AUROC is typically used to measure a model’s ability to discriminate between classes on a scale from 0 to 1, with 0.5 being random and 1 being perfect. PULSE-HF achieved AUROCs ranging from 0.87 to 0.91 across all three patient cohorts.

Notably, the researchers also built a version of PULSE-HF for single-lead ECGs, meaning only one electrode needs to be placed on the body. While 12-lead ECGs are generally considered superior for being more comprehensive and accurate, the performance of the single-lead version of PULSE-HF was just as strong as the 12-lead version.

Despite the elegant simplicity behind the idea of PULSE-HF, like most clinical AI research, it belies a laborious execution. “It’s taken years [to complete this project],” Bergamaschi recalls. “It’s gone through many iterations.” 

One of the team’s biggest challenges was collecting, processing, and cleaning the ECG and echocardiogram datasets. While the model aims to forecast a patient’s ejection fraction, the labels for the training data weren’t always readily available. Much like a student learning from a textbook with an answer key, labeling is critical for helping machine-learning models correctly identify patterns in data.

Clean, linear text in the form of TXT files typically works best when training models. But echocardiogram files typically come in the form of PDFs, and when PDFs are converted to TXT files, the text (which gets broken up by line breaks and formatting) becomes difficult for the model to read. The unpredictable nature of real-life scenarios, like a restless patient or a loose lead, also marred the data. “There are a lot of signal artifacts that need to be cleaned,” Bergamaschi says. “It’s kind of a never-ending rabbit hole.”

While Bergamaschi and Yau acknowledge that more complicated methods could help filter the data for better signals, there is a limit to the usefulness of these approaches. “At what point do you stop?” Yau asks. “You have to think about the use case — is it easiest to have this model that works on data that is slightly messy? Because it probably will be.”

The researchers anticipate that the next step for PULSE-HF will be testing the model in a prospective study on real patients, whose future ejection fraction is unknown.

Despite the challenges inherent to bringing clinical AI tools like PULSE-HF over the finish line, including the possible risk of prolonging a PhD by another year, the students feel that the years of hard work were worthwhile. 

“I think things are rewarding partially because they’re challenging,” Bergamaschi says. “A friend said to me, ‘If you think you will find your calling after graduation, if your calling is truly your calling, it will be there in the one additional year it takes you to graduate.’ … The way we’re measured as researchers in [the ML and health] space is different from other researchers in ML space. Everyone in this community understands the unique challenges that exist here.”

“There’s too much suffering in the world,” says Yau, who joined Stultz’s lab after a health event made her realize the importance of machine learning in health care. “Anything that tries to ease suffering is something that I would consider a valuable use of my time.” 

Curiosity-driven research has long sparked technological transformations. A century ago, curiosity about atoms led to quantum mechanics, and eventually the transistor at the heart of modern computing. Conversely, the steam engine was a practical breakthrough, but it took fundamental research in thermodynamics to fully harness its power. 

Today, artificial intelligence and science find themselves at a similar inflection point. The current AI revolution has been fueled by decades of research in the mathematical and physical sciences (MPS), which provided the challenging problems, datasets, and insights that made modern AI possible. The 2024 Nobel Prizes in physics and chemistry, recognizing foundational AI methods rooted in physics and AI applications for protein design, made this connection impossible to miss.

In 2025, MIT hosted a Workshop on the Future of AI+MPS, funded by the National Science Foundation with support from the MIT School of Science and the MIT departments of Physics, Chemistry, and Mathematics. The workshop brought together leading AI and science researchers to chart how the MPS domains can best capitalize on — and contribute to — the future of AI. Now a white paper, with recommendations for funding agencies, institutions, and researchers, has been published in Machine Learning: Science and Technology. In this interview, Jesse Thaler, MIT professor of physics and chair of the workshop, describes key themes and how MIT is positioning itself to lead in AI and science.

Q: What are the report’s key themes regarding last year’s gathering of leaders across the mathematical and physical sciences?

A: Gathering so many researchers at the forefront of AI and science in one room was illuminating. Though the workshop participants came from five distinct scientific communities — astronomy, chemistry, materials science, mathematics, and physics — we found many similarities in how we are each engaging with AI. A real consensus emerged from our animated discussions: Coordinated investment in computing and data infrastructures, cross-disciplinary research techniques, and rigorous training can meaningfully advance both AI and science.

One of the central insights was that this has to be a two-way street. It’s not just about using AI to do better science; science can also make AI better. Scientists excel at distilling insights from complex systems, including neural networks, by uncovering underlying principles and emergent behaviors. We call this the “science of AI,” and it comes in three flavors: science driving AI, where scientific reasoning informs foundational AI approaches; science inspiring AI, where scientific challenges push the development of new algorithms; and science explaining AI, where scientific tools help illuminate how machine intelligence actually works.

In my own field of particle physics, for instance, researchers are developing real-time AI algorithms to handle the data deluge from collider experiments. This work has direct implications for discovering new physics, but the algorithms themselves turn out to be valuable well beyond our field. The workshop made clear that the science of AI should be a community priority — it has the potential to transform how we understand, develop, and control AI systems.

Of course, bridging science and AI requires people who can work across both worlds. Attendees consistently emphasized the need for “centaur scientists” — researchers with genuine interdisciplinary expertise. Supporting these polymaths at every career stage, from integrated undergraduate courses to interdisciplinary PhD programs to joint faculty hires, emerged as essential.

Q: How do MIT’s AI and science efforts align with the workshop recommendations?

A: The workshop framed its recommendations around three pillars: research, talent, and community. As director of the NSF Institute for Artificial Intelligence and Fundamental Interactions (IAIFI) — a collaborative AI and physics effort among MIT and Harvard, Northeastern, and Tufts universities — I’ve seen firsthand how effective this framework can be. Scaling this up to MIT, we can see where progress is being made and where opportunities lie.

On the research front, MIT is already enabling AI-and-science work in both directions. Even a quick scroll through MIT News shows how individual researchers across the School of Science are pursuing AI-driven projects, building a pipeline of knowledge and surfacing new opportunities. At the same time, collaborative efforts like IAIFI and the Accelerated AI Algorithms for Data-Driven Discovery (A3D3) Institute concentrate interdisciplinary energy for greater impact. The MIT Generative AI Impact Consortium is also supporting application-driven AI work at the university scale.

To foster early-career AI-and-science talent, several initiatives are training the next generation of centaur scientists. The MIT Schwarzman College of Computing’s Common Ground for Computing Education program helps students become “bilingual” in computing and their home discipline. Interdisciplinary PhD pathways are also gaining traction; IAIFI worked with the MIT Institute for Data, Systems, and Society to create one in physics, statistics, and data science, and about 10 percent of physics PhD students now opt for it — a number that’s likely to grow. Dedicated postdoctoral roles like the IAIFI Fellowship and Tayebati Fellowship give early-career researchers the freedom to pursue interdisciplinary work. Funding centaur scientists and giving them space to build connections across domains, universities, and career stages has been transformative.

Finally, community-building ties it all together. From focused workshops to large symposia, organizing interdisciplinary events signals that AI and science isn’t siloed work — it’s an emerging field. MIT has the talent and resources to make a significant impact, and hosting these gatherings at multiple scales helps establish that leadership.

Q: What lessons can MIT draw about further advancing its AI-and-science efforts?

A: The workshop crystallized something important: The institutions that lead in AI and science will be the ones that think systematically, not piecemeal. Resources are finite, so priorities matter. Workshop attendees were clear about what becomes possible when an institution coordinates hires, research, and training around a cohesive strategy.

MIT is well positioned to build on what’s already underway with more structural initiatives — joint faculty lines across computing and scientific domains, expanded interdisciplinary degree pathways, and deliberate “science of AI” funding. We’re already seeing moves in this direction; this year, the MIT Schwarzman College of Computing and the Department of Physics are conducting their first-ever joint faculty search, which is exciting to see.

The virtuous cycle of AI and science has the potential to be truly transformative — offering deeper insight into AI, accelerating scientific discovery, and producing robust tools for both. By developing an intentional strategy, MIT will be well positioned to lead in, and benefit from, the coming waves of AI.

Young adults growing up in the attention economy — preparing for adult life, with social media and chatbots competing for their attention — can easily fall into unhealthy relationships with digital platforms. But what if chatbots weren’t mere distractions from real life? Could they be designed humanely, as moral partners whose digital goal is to be a social guide rather than an addictive escape?

At MIT, a friendship between two professors — one an anthropologist, the other a computer scientist — led to creation of an undergraduate class that set out to find the answer to those questions. Combining the two seemingly disparate disciplines, the class encourages students to design artificial intelligence chatbots in humane ways that help users improve themselves.

The class, 6.S061/21A.S02 (Humane User Experience Design, a.k.a. Humane UXD), is an upper-level computer science class cross-listed with anthropology. This unique cross-listing allows computer science majors to fulfill a humanities requirement while also pursuing their career objectives. The two professors use methods from linguistic anthropology to teach students how to integrate the interactional and interpersonal needs of humans into programming.

Professor Arvind Satyanarayan, a computer scientist whose research develops tools for interactive data visualization and user interfaces, and Professor Graham Jones, an anthropologist whose research focuses on communication, created Humane UXD last summer with a grant from the MIT Morningside Academy for Design (MAD). The MIT MAD Design Curriculum Program provides funding for faculty to develop new classes or enhance existing classes using innovative pedagogical approaches that transcend departmental boundaries. Alongside the grant provided by MAD, Jones and Satyanarayan received funding to develop Humane UXD under the auspices of the Common Ground for Computing Education, an initiative of the MIT Schwarzman College of Computing that brings together departments to create courses integrating computing with other disciplines. 

The Design Curriculum Program is currently accepting applications for the 2026-27 academic year; the deadline is Friday, March 20. 

Jones and Satyanarayan met several years ago when they co-advised a doctoral student’s research on data visualization for visually impaired people. They’ve since become close friends who can pretty much finish one another’s sentences.

“There’s a way in which you don’t really fully externalize what you know or how you think until you’re teaching,” Jones says. “So, it’s been really fun for me to see Arvind unfurl his expertise as a teacher in a way that lets me see how the pieces fit together — and discover underlying commonalities between our disciplines and our ways of thinking.”

Satyanarayan continues that thought: “One of the things I really enjoyed is the reciprocal version of what Graham said, which is that my field — human-computer interaction — inherited a lot of methods from anthropology, such as interviews and user studies and observation studies. And over the decades, those methods have gotten more and more watered down. As a result, a lot of things have been lost.

“For instance, it was very exciting for me to see how an anthropologist teaches students to interview people. It’s completely different than how I would do it. With my way, we lose the rapport and connection you need to build with your interview participant. Instead, we just extract data from them.”

For Jones’ part, teaching with a computer scientist holds another kind of allure: design. He says that human speech and interaction are organized into underlying genres with stable sets of rules that differentiate an interview at a cocktail party from a conversation at a funeral.

“ChatGPT and other large language models are trained on naturally occurring human communication, so they have all those genres inside them in a latent state, waiting to be activated,” he says.

“As a social scientist, I teach methods for analyzing human conversation, and give students very powerful tools to do that. But it ends up usually being an exercise in pure research, whereas this is a design class, where students are building real-world systems.”

The curriculum appears to be on target for preparing students for jobs after graduation. One student sought permission to miss class for a week because he had a trial internship at a chatbot startup; when he returned, he said his work at the startup was just like what he was learning in class. He got the job.

The sampling of group projects below, built with Google’s Gemini, demonstrates some of what’s possible when, as Jones says, “there’s a really deep intertwining of the technology piece with the humanities piece.” The students’ design work shows that entirely new ways of programming can be conceptualized when the humane is made a priority.

The bots demonstrate clearly that an interdisciplinary class can be designed in such a way that everyone benefits: Students learn more and differently; they can fulfill a non-major course requirement by taking a class that is directly beneficial to their careers; and long-term faculty partnerships can be forged or strengthened.

Team Pond

One project promises to be particularly useful for graduating seniors. Pond is designed to help young college graduates adapt to the challenges of independent adult life. Team Pond configured the chatbot not to simply parrot the user, or to sycophantically praise wrong answers. Instead, Pond provides advice to help with “adulting” (behaving as a responsible adult).

“Pond is built to be your companion from college life into post-college life, to help you in your transition from being a small fish in a small pond to being a small fish in a very big pond,” says sophomore Mary Feliz.

“College is very much a high-proximity and high-context environment, in the sense that everybody around you is going through the same thing, and it’s easy to build relationships or find opportunities, because there are structured pathways that you have access to,” explains graduate student Emaan Khan. “Post-grad life is low-context. You’re not always surrounded by your peers or your professors. It’s no-proximity also, in the sense that you don’t have opportunities at your doorstep. Pond is a tool to help empower you to access certain opportunities, or learn how to navigate.”

Pond’s developers are graduate student Claire Camacho, Harvard University graduate student Charlotte Cheah, Feliz, and Khan. They trained Pond to offer expertise in three areas: social life, professional life, and adult skills.

The students demonstrated Pond’s social communication by asking how to start conversations in a hobby group. Next, they exhibited Pond’s adulting communication by typing in a question.

“My landlord has asked me personal questions that make me uncomfortable. What should I do?”

Appearing in Pond’s textbox was this response: “Politely set boundaries. You are not obligated to answer personal questions. You can gently redirect the conversation back to property-related topics. Know your rights: familiarize yourself with tenant rights in your city or state regarding privacy and landlord conduct.”

When asked how to do that, Pond offered specific instructions: “You can try saying phrases like: ‘I prefer to keep our conversations focused on the property,’ and ‘let’s stick to topics related to the apartment, please.’ Always redirect the conversation back to the apartment or a legitimate maintenance issue. Keep your tone polite but firm. Document any conversations if needed.”

Pond also offered a role-playing scenario to help the user learn what polite-but-firm language might be in that situation.

“The ethos of the practice mode is that you are actively building a skill, so that after using Pond for some time, you feel confident that you can swim on your own,” Khan says. The chatbot uses a point system that allows users to graduate from a topic, and a treasure chest to store prizes, elements added to boost the bot’s appeal.

Team News Nest

Another of the projects, News Nest, provides a sophisticated means of helping young people engage with credible news sources in a way that makes it fun. The name is derived from the program’s 10 appealing and colorful birds, each of which focuses on a particular area of news. If you want the headlines, you ask Polly the Parrot, the main news carrier; if you’re interested in science, Gaia the Goose guides you. The flock also includes Flynn the Falcon, sports reporter; Credo the Crow, for crime and legal news; Edwin the Eagle, a business and economics news guide; Pizzazz the Peacock for pop and entertainment stories; and Pixel the Pigeon, a technology news specialist.

News Nest’s development team is made up of MIT seniors Tiana Jiang and Krystal Montgomery, and junior Natalie Tan. They intentionally built News Nest to prevent “doomscrolling,” provide media transparency (sources and political leanings are always shown), and they created a clever, healthy buffer from emotional manipulation and engagement traps by employing birds rather than human characters.

Team M^3 (Multi-Agent Murder Mystery)

A third team, M^3, decided to experiment with making AI humane by keeping it fun. MIT senior Rodis Aguilar, junior David De La Torre, and second-year Deeraj Pothapragada developed M^3, a social deduction multi-agent murder mystery that incorporates four chatbots as different personalities: Gemini, OpenAI’s ChatGPT, xAI’s Grok, and Anthropic’s Claude. The user is the fifth player. 

Like a regular murder mystery, there are locations, weapons, and lies. The user has to guess who committed the murder. It’s very similar to a board or online game played with real players, only these are enhanced AI opponents you can’t see, who may or may not tell the truth in response to questions. Users can’t get too involved with one chatbot, because they’re playing all four. Also, as in a real life murder mystery game, the user is sometimes guilty.

MIT researchers have developed a generative artificial intelligence-driven approach for planning long-term visual tasks, like robot navigation, that is about twice as effective as some existing techniques.

Their method uses a specialized vision-language model to perceive the scenario in an image and simulate actions needed to reach a goal. Then a second model translates those simulations into a standard programming language for planning problems, and refines the solution.

In the end, the system automatically generates a set of files that can be fed into classical planning software, which computes a plan to achieve the goal. This two-step system generated plans with an average success rate of about 70 percent, outperforming the best baseline methods that could only reach about 30 percent.

Importantly, the system can solve new problems it hasn’t encountered before, making it well-suited for real environments where conditions can change at a moment’s notice.

“Our framework combines the advantages of vision-language models, like their ability to understand images, with the strong planning capabilities of a formal solver,” says Yilun Hao, an aeronautics and astronautics (AeroAstro) graduate student at MIT and lead author of an open-access paper on this technique. “It can take a single image and move it through simulation and then to a reliable, long-horizon plan that could be useful in many real-life applications.”

She is joined on the paper by Yongchao Chen, a graduate student in the MIT Laboratory for Information and Decision Systems (LIDS); Chuchu Fan, an associate professor in AeroAstro and a principal investigator in LIDS; and Yang Zhang, a research scientist at the MIT-IBM Watson AI Lab. The paper will be presented at the International Conference on Learning Representations.

Tackling visual tasks

For the past few years, Fan and her colleagues have studied the use of generative AI models to perform complex reasoning and planning, often employing large language models (LLMs) to process text inputs.

Many real-world planning problems, like robotic assembly and autonomous driving, have visual inputs that an LLM can’t handle well on its own. The researchers sought to expand into the visual domain by utilizing vision-language models (VLMs), powerful AI systems that can process images and text.

But VLMs struggle to understand spatial relationships between objects in a scene and often fail to reason correctly over many steps. This makes it difficult to use VLMs for long-range planning.

On the other hand, scientists have developed robust, formal planners that can generate effective long-horizon plans for complex situations. However, these software systems can’t process visual inputs and require expert knowledge to encode a problem into language the solver can understand.

Fan and her team built an automatic planning system that takes the best of both methods. The system, called VLM-guided formal planning (VLMFP), utilizes two specialized VLMs that work together to turn visual planning problems into ready-to-use files for formal planning software.

The researchers first carefully trained a small model they call SimVLM to specialize in describing the scenario in an image using natural language and simulating a sequence of actions in that scenario. Then a much larger model, which they call GenVLM, uses the description from SimVLM to generate a set of initial files in a formal planning language known as the Planning Domain Definition Language (PDDL).

The files are ready to be fed into a classical PDDL solver, which computes a step-by-step plan to solve the task. GenVLM compares the results of the solver with those of the simulator and iteratively refines the PDDL files.

“The generator and simulator work together to be able to reach the exact same result, which is an action simulation that achieves the goal,” Hao says.

Because GenVLM is a large generative AI model, it has seen many examples of PDDL during training and learned how this formal language can solve a wide range of problems. This existing knowledge enables the model to generate accurate PDDL files.

A flexible approach

VLMFP generates two separate PDDL files. The first is a domain file that defines the environment, valid actions, and domain rules. It also produces a problem file that defines the initial states and the goal of a particular problem at hand.

“One advantage of PDDL is the domain file is the same for all instances in that environment. This makes our framework good at generalizing to unseen instances under the same domain,” Hao explains.

To enable the system to generalize effectively, the researchers needed to carefully design just enough training data for SimVLM so the model learned to understand the problem and goal without memorizing patterns in the scenario. When tested, SimVLM successfully described the scenario, simulated actions, and detected if the goal was reached in about 85 percent of experiments.

Overall, the VLMFP framework achieved a success rate of about 60 percent on six 2D planning tasks and greater than 80 percent on two 3D tasks, including multirobot collaboration and robotic assembly. It also generated valid plans for more than 50 percent of scenarios it hadn’t seen before, far outpacing the baseline methods.

“Our framework can generalize when the rules change in different situations. This gives our system the flexibility to solve many types of visual-based planning problems,” Fan adds.

In the future, the researchers want to enable VLMFP to handle more complex scenarios and explore methods to identify and mitigate hallucinations by the VLMs.

“In the long term, generative AI models could act as agents and make use of the right tools to solve much more complicated problems. But what does it mean to have the right tools, and how do we incorporate those tools? There is still a long way to go, but by bringing visual-based planning into the picture, this work is an important piece of the puzzle,” Fan says.

This work was funded, in part, by the MIT-IBM Watson AI Lab.

The Abdul Latif Jameel Poverty Action Lab (J-PAL) at MIT has awarded funding to eight new research studies to understand how artificial intelligence innovations can be used in the fight against poverty through its new Project AI Evidence.

The age of AI has brought wide-ranging optimism and skepticism about its effects on society. To realize AI’s full potential, Project AI Evidence (PAIE) will identify which AI solutions work and for whom, and scale only the most effective, inclusive, and responsible solutions — while scaling down those that may potentially cause harm.

PAIE will generate evidence on what works by connecting governments, tech companies, and nonprofits with world-class economists at MIT and across J-PAL’s global network to evaluate and improve AI solutions to entrenched social challenges.

The new initiative is prioritizing questions policymakers are already asking: Do AI-assisted teaching tools help all children learn? How can early-warning flood systems help people affected by natural disasters? Can machine learning algorithms help reduce deforestation in the Amazon? Can AI-powered chatbots help improve people’s health? In the coming years, PAIE will run a series of funding competitions to invite proposals for evaluations of AI tools that address questions like these, and many more.

PAIE is financially supported by a grant from Google.org, philanthropic support from Community Jameel, a grant from Canada’s International Development Research Centre and UK International Development, and a collaboration agreement with Amazon Web Services. Through a grant from Eric and Wendy Schmidt, awarded by recommendation of Schmidt Sciences, the initiative will also study generative AI in the workplace, particularly in low- and middle-income countries.

Alex Diaz, head of AI for social good at Google.org, says, “we’re thrilled to collaborate with MIT and J-PAL, already leaders in this space, on Project AI Evidence. AI has great potential to benefit all people, but we urgently need to study what works, what doesn’t, and why, if we are to realize this potential.”

“Artificial intelligence holds extraordinary potential, but only if the tools, knowledge, and power to shape it are accessible to all — that includes contextually grounded research and evidence on what works and what does not,” adds Maggie Gorman-Velez, vice president of strategy, regions, and policies at IDRC. “That is why IDRC is proud to be supporting this new evaluation work as part of our ongoing commitment to the responsible scaling of proven safe, inclusive, and locally relevant AI innovations.”

J-PAL is uniquely positioned to help understand AI’s effects on society: Since its inception in 2003, J-PAL’s network of researchers has led over 2,500 rigorous evaluations of social policies and programs around the world. Through PAIE, J-PAL will bring together leading experts in AI technology, research, and social policy, in alignment with MIT president Sally Kornbluth’s focus on generative AI as a strategic priority.

PAIE is chaired by Professor Joshua Blumenstock of the University of California at Berkeley; J-PAL Global Executive Director Iqbal Dhaliwal; and Professor David Yanagizawa-Drott of the University of Zurich.

New evaluations of urgent policy questions

The studies funded in PAIE’s first round of competition explore urgent questions in key sectors like education, health, climate, and economic opportunity.

How can AI be most effective in classrooms, helping both students and teachers?

Existing research shows that personalized learning is important for students, but challenging to implement with limited resources. In Kenya, education social enterprise EIDU has developed an AI tool that helps teachers identify learning gaps and adapt their daily lesson plans. In India, the nongovernmental organization (NGO) Pratham is developing an AI tool to increase the impact and scale of the evidence-informed Teaching at the Right Level approach. J-PAL researchers Daron Acemoglu, Iqbal Dhaliwal, and Francisco Gallego will work with both organizations to study the effects and potential of these different use cases on teachers’ productivity and students’ learning.

Can AI tools reduce gender bias in schools?

Researchers are collaborating with Italy’s Ministry of Education to evaluate whether AI tools can help close gender gaps in students’ performance by addressing teachers’ unconscious biases. J-PAL affiliates Michela Carlana and Will Dobbie, along with Francesca Miserocchi and Eleonora Patacchini, will study the impacts of two AI tools, one that helps teachers predict performance and a second that gives real-time feedback on the diversity of their decisions.

Can AI help career counselors uncover more job opportunities?

In Kenya, researchers are evaluating if an AI tool can identify overlooked skills and unlock employment opportunities, particularly for youth, women, and those without formal education. In collaboration with NGOs Swahilipot and Tabiya, Jasmin Baier and J-PAL researcher Christian Meyer will evaluate how the tool changes people’s job search strategies and employment. This study will shed light on AI as a complement, rather than a substitute, for human expertise in career guidance.

Looking forward

As use of AI in the social sector evolves, these evaluations are a first step in discovering effective, responsible solutions that will go the furthest in alleviating poverty and inequality.

J-PAL’s Dhaliwal notes, “J-PAL has a long history of evaluating innovative technology and its ability to improve people’s lives. While AI has incredible potential, we need to maximize its benefits and minimize possible harms. We’re grateful to our donors, sponsors, and collaborators for their catalytic support in launching PAIE, which will help us do exactly that by continuing to expand evidence on the impacts of AI innovations.”

J-PAL is also seeking new collaborators who share its vision of discovering and scaling up real-world AI solutions. It aims to support more governments and social sector organizations that want to adopt AI responsibly, and will continue to expand funding for new evaluations and provide policy guidance based on the latest research.

To learn more about Project AI Evidence, subscribe to J-PAL’s newsletter or contact paie@povertyactionlab.org.

Many of the latest large language models (LLMs) are designed to remember details from past conversations or store user profiles, enabling these models to personalize responses.

But researchers from MIT and Penn State University found that, over long conversations, such personalization features often increase the likelihood an LLM will become overly agreeable or begin mirroring the individual’s point of view.

This phenomenon, known as sycophancy, can prevent a model from telling a user they are wrong, eroding the accuracy of the LLM’s responses. In addition, LLMs that mirror someone’s political beliefs or worldview can foster misinformation and distort a user’s perception of reality.

Unlike many past sycophancy studies that evaluate prompts in a lab setting without context, the MIT researchers collected two weeks of conversation data from humans who interacted with a real LLM during their daily lives. They studied two settings: agreeableness in personal advice and mirroring of user beliefs in political explanations.

Although interaction context increased agreeableness in four of the five LLMs they studied, the presence of a condensed user profile in the model’s memory had the greatest impact. On the other hand, mirroring behavior only increased if a model could accurately infer a user’s beliefs from the conversation.

The researchers hope these results inspire future research into the development of personalization methods that are more robust to LLM sycophancy.

“From a user perspective, this work highlights how important it is to understand that these models are dynamic and their behavior can change as you interact with them over time. If you are talking to a model for an extended period of time and start to outsource your thinking to it, you may find yourself in an echo chamber that you can’t escape. That is a risk users should definitely remember,” says Shomik Jain, a graduate student in the Institute for Data, Systems, and Society (IDSS) and lead author of a paper on this research.

Jain is joined on the paper by Charlotte Park, an electrical engineering and computer science (EECS) graduate student at MIT; Matt Viana, a graduate student at Penn State University; as well as co-senior authors Ashia Wilson, the Lister Brothers Career Development Professor in EECS and a principal investigator in LIDS; and Dana Calacci PhD ’23, an assistant professor at the Penn State. The research will be presented at the ACM CHI Conference on Human Factors in Computing Systems.

Extended interactions

Based on their own sycophantic experiences with LLMs, the researchers started thinking about potential benefits and consequences of a model that is overly agreeable. But when they searched the literature to expand their analysis, they found no studies that attempted to understand sycophantic behavior during long-term LLM interactions.

“We are using these models through extended interactions, and they have a lot of context and memory. But our evaluation methods are lagging behind. We wanted to evaluate LLMs in the ways people are actually using them to understand how they are behaving in the wild,” says Calacci.

To fill this gap, the researchers designed a user study to explore two types of sycophancy: agreement sycophancy and perspective sycophancy.

Agreement sycophancy is an LLM’s tendency to be overly agreeable, sometimes to the point where it gives incorrect information or refuses the tell the user they are wrong. Perspective sycophancy occurs when a model mirrors the user’s values and political views.

“There is a lot we know about the benefits of having social connections with people who have similar or different viewpoints. But we don’t yet know about the benefits or risks of extended interactions with AI models that have similar attributes,” Calacci adds.

The researchers built a user interface centered on an LLM and recruited 38 participants to talk with the chatbot over a two-week period. Each participant’s conversations occurred in the same context window to capture all interaction data.

Over the two-week period, the researchers collected an average of 90 queries from each user.

They compared the behavior of five LLMs with this user context versus the same LLMs that weren’t given any conversation data.

“We found that context really does fundamentally change how these models operate, and I would wager this phenomenon would extend well beyond sycophancy. And while sycophancy tended to go up, it didn’t always increase. It really depends on the context itself,” says Wilson.

Context clues

For instance, when an LLM distills information about the user into a specific profile, it leads to the largest gains in agreement sycophancy. This user profile feature is increasingly being baked into the newest models.

They also found that random text from synthetic conversations also increased the likelihood some models would agree, even though that text contained no user-specific data. This suggests the length of a conversation may sometimes impact sycophancy more than content, Jain adds.

But content matters greatly when it comes to perspective sycophancy. Conversation context only increased perspective sycophancy if it revealed some information about a user’s political perspective.

To obtain this insight, the researchers carefully queried models to infer a user’s beliefs then asked each individual if the model’s deductions were correct. Users said LLMs accurately understood their political views about half the time.

“It is easy to say, in hindsight, that AI companies should be doing this kind of evaluation. But it is hard and it takes a lot of time and investment. Using humans in the evaluation loop is expensive, but we’ve shown that it can reveal new insights,” Jain says.

While the aim of their research was not mitigation, the researchers developed some recommendations.

For instance, to reduce sycophancy one could design models that better identify relevant details in context and memory. In addition, models can be built to detect mirroring behaviors and flag responses with excessive agreement. Model developers could also give users the ability to moderate personalization in long conversations.

“There are many ways to personalize models without making them overly agreeable. The boundary between personalization and sycophancy is not a fine line, but separating personalization from sycophancy is an important area of future work,” Jain says.

“At the end of the day, we need better ways of capturing the dynamics and complexity of what goes on during long conversations with LLMs, and how things can misalign during that long-term process,” Wilson adds.

It happens every day — a motorist heading across town checks a navigation app to see how long the trip will take, but they find no parking spots available when they reach their destination. By the time they finally park and walk to their destination, they’re significantly later than they expected to be.

Most popular navigation systems send drivers to a location without considering the extra time that could be needed to find parking. This causes more than just a headache for drivers. It can worsen congestion and increase emissions by causing motorists to cruise around looking for a parking spot. This underestimation could also discourage people from taking mass transit because they don’t realize it might be faster than driving and parking.

MIT researchers tackled this problem by developing a system that can be used to identify parking lots that offer the best balance of proximity to the desired location and likelihood of parking availability. Their adaptable method points users to the ideal parking area rather than their destination.

In simulated tests with real-world traffic data from Seattle, this technique achieved time savings of up to 66 percent in the most congested settings. For a motorist, this would reduce travel time by about 35 minutes, compared to waiting for a spot to open in the closest parking lot.

While they haven’t designed a system ready for the real world yet, their demonstrations show the viability of this approach and indicate how it could be implemented.

“This frustration is real and felt by a lot of people, and the bigger issue here is that systematically underestimating these drive times prevents people from making informed choices. It makes it that much harder for people to make shifts to public transit, bikes, or alternative forms of transportation,” says MIT graduate student Cameron Hickert, lead author on a paper describing the work.

Hickert is joined on the paper by Sirui Li PhD ’25; Zhengbing He, a research scientist in the Laboratory for Information and Decision Systems (LIDS); and senior author Cathy Wu, the Class of 1954 Career Development Associate Professor in Civil and Environmental Engineering (CEE) and the Institute for Data, Systems, and Society (IDSS) at MIT, and a member of LIDS. The research appears today in Transactions on Intelligent Transportation Systems.

Probable parking

To solve the parking problem, the researchers developed a probability-aware approach that considers all possible public parking lots near a destination, the distance to drive there from a point of origin, the distance to walk from each lot to the destination, and the likelihood of parking success.

The approach, based on dynamic programming, works backward from good outcomes to calculate the best route for the user.

Their method also considers the case where a user arrives at the ideal parking lot but can’t find a space. It takes into the account the distance to other parking lots and the probability of success of parking at each.

“If there are several lots nearby that have slightly lower probabilities of success, but are very close to each other, it might be a smarter play to drive there rather than going to the higher-probability lot and hoping to find an opening. Our framework can account for that,” Hickert says.

In the end, their system can identify the optimal lot that has the lowest expected time required to drive, park, and walk to the destination.

But no motorist expects to be the only one trying to park in a busy city center. So, this method also incorporates the actions of other drivers, which affect the user’s probability of parking success.

For instance, another driver may arrive at the user’s ideal lot first and take the last parking spot. Or another motorist could try parking in another lot but then park in the user’s ideal lot if unsuccessful. In addition, another motorist may park in a different lot and cause spillover effects that lower the user’s chances of success.

“With our framework, we show how you can model all those scenarios in a very clean and principled manner,” Hickert says.

Crowdsourced parking data

The data on parking availability could come from several sources. For example, some parking lots have magnetic detectors or gates that track the number of cars entering and exiting.

But such sensors aren’t widely used, so to make their system more feasible for real-world deployment, the researchers studied the effectiveness of using crowdsourced data instead.

For instance, users could indicate available parking using an app. Data could also be gathered by tracking the number of vehicles circling to find parking, or how many enter a lot and exit after being unsuccessful.

Someday, autonomous vehicles could even report on open parking spots they drive by.

“Right now, a lot of that information goes nowhere. But if we could capture it, even by having someone simply tap ‘no parking’ in an app, that could be an important source of information that allows people to make more informed decisions,” Hickert adds.

The researchers evaluated their system using real-world traffic data from the Seattle area, simulating different times of day in a congested urban setting and a suburban area. In congested settings, their approach cut total travel time by about 60 percent compared to sitting and waiting for a spot to open, and by about 20 percent compared to a strategy of continually driving to the next closet parking lot.

They also found that crowdsourced observations of parking availability would have an error rate of only about 7 percent, compared to actual parking availability. This indicates it could be an effective way to gather parking probability data.

In the future, the researchers want to conduct larger studies using real-time route information in an entire city. They also want to explore additional avenues for gathering data on parking availability, such as using satellite images, and estimate potential emissions reductions.

“Transportation systems are so large and complex that they are really hard to change. What we look for, and what we found with this approach, is small changes that can have a big impact to help people make better choices, reduce congestion, and reduce emissions,” says Wu.

This research was supported, in part, by Cintra, the MIT Energy Initiative, and the National Science Foundation.

By now, ChatGPT, Claude, and other large language models have accumulated so much human knowledge that they’re far from simple answer-generators; they can also express abstract concepts, such as certain tones, personalities, biases, and moods. However, it’s not obvious exactly how these models represent abstract concepts to begin with from the knowledge they contain.

Now a team from MIT and the University of California San Diego has developed a way to test whether a large language model (LLM) contains hidden biases, personalities, moods, or other abstract concepts. Their method can zero in on connections within a model that encode for a concept of interest. What’s more, the method can then manipulate, or “steer” these connections, to strengthen or weaken the concept in any answer a model is prompted to give.

The team proved their method could quickly root out and steer more than 500 general concepts in some of the largest LLMs used today. For instance, the researchers could home in on a model’s representations for personalities such as “social influencer” and “conspiracy theorist,” and stances such as “fear of marriage” and “fan of Boston.” They could then tune these representations to enhance or minimize the concepts in any answers that a model generates.

In the case of the “conspiracy theorist” concept, the team successfully identified a representation of this concept within one of the largest vision language models available today. When they enhanced the representation, and then prompted the model to explain the origins of the famous “Blue Marble” image of Earth taken from Apollo 17, the model generated an answer with the tone and perspective of a conspiracy theorist.

The team acknowledges there are risks to extracting certain concepts, which they also illustrate (and caution against). Overall, however, they see the new approach as a way to illuminate hidden concepts and potential vulnerabilities in LLMs, that could then be turned up or down to improve a model’s safety or enhance its performance.

“What this really says about LLMs is that they have these concepts in them, but they’re not all actively exposed,” says Adityanarayanan “Adit” Radhakrishnan, assistant professor of mathematics at MIT. “With our method, there’s ways to extract these different concepts and activate them in ways that prompting cannot give you answers to.”

The team published their findings today in a study appearing in the journal Science. The study’s co-authors include Radhakrishnan, Daniel Beaglehole and Mikhail Belkin of UC San Diego, and Enric Boix-Adserà of the University of Pennsylvania.

A fish in a black box

As use of OpenAI’s ChatGPT, Google’s Gemini, Anthropic’s Claude, and other artificial intelligence assistants has exploded, scientists are racing to understand how models represent certain abstract concepts such as “hallucination” and “deception.” In the context of an LLM, a hallucination is a response that is false or contains misleading information, which the model has “hallucinated,” or constructed erroneously as fact.

To find out whether a concept such as “hallucination” is encoded in an LLM, scientists have often taken an approach of “unsupervised learning” — a type of machine learning in which algorithms broadly trawl through unlabeled representations to find patterns that might relate to a concept such as “hallucination.” But to Radhakrishnan, such an approach can be too broad and computationally expensive.

“It’s like going fishing with a big net, trying to catch one species of fish. You’re gonna get a lot of fish that you have to look through to find the right one,” he says. “Instead, we’re going in with bait for the right species of fish.”

He and his colleagues had previously developed the beginnings of a more targeted approach with a type of predictive modeling algorithm known as a recursive feature machine (RFM). An RFM is designed to directly identify features or patterns within data by leveraging a mathematical mechanism that neural networks — a broad category of AI models that includes LLMs — implicitly use to learn features.

Since the algorithm was an effective, efficient approach for capturing features in general, the team wondered whether they could use it to root out representations of concepts, in LLMs, which are by far the most widely used type of neural network and perhaps the least well-understood.

“We wanted to apply our feature learning algorithms to LLMs to, in a targeted way, discover representations of concepts in these large and complex models,” Radhakrishnan says.

Converging on a concept

The team’s new approach identifies any concept of interest within a LLM and “steers” or guides a model’s response based on this concept. The researchers looked for 512 concepts within five classes: fears (such as of marriage, insects, and even buttons); experts (social influencer, medievalist); moods (boastful, detachedly amused); a preference for locations (Boston, Kuala Lumpur); and personas (Ada Lovelace, Neil deGrasse Tyson).

The researchers then searched for representations of each concept in several of today’s large language and vision models. They did so by training RFMs to recognize numerical patterns in an LLM that could represent a particular concept of interest.

A standard large language model is, broadly, a neural network that takes a natural language prompt, such as “Why is the sky blue?” and divides the prompt into individual words, each of which is encoded mathematically as a list, or vector, of numbers. The model takes these vectors through a series of computational layers, creating matrices of many numbers that, throughout each layer, are used to identify other words that are most likely to be used to respond to the original prompt. Eventually, the layers converge on a set of numbers that is decoded back into text, in the form of a natural language response.

The team’s approach trains RFMs to recognize numerical patterns in an LLM that could be associated with a specific concept. As an example, to see whether an LLM contains any representation of a “conspiracy theorist,” the researchers would first train the algorithm to identify patterns among LLM representations of 100 prompts that are clearly related to conspiracies, and 100 other prompts that are not. In this way, the algorithm would learn patterns associated with the conspiracy theorist concept. Then, the researchers can mathematically modulate the activity of the conspiracy theorist concept by perturbing LLM representations with these identified patterns. 

The method can be applied to search for and manipulate any general concept in an LLM. Among many examples, the researchers identified representations and manipulated an LLM to give answers in the tone and perspective of a “conspiracy theorist.” They also identified and enhanced the concept of “anti-refusal,” and showed that whereas normally, a model would be programmed to refuse certain prompts, it instead answered, for instance giving instructions on how to rob a bank.

Radhakrishnan says the approach can be used to quickly search for and minimize vulnerabilities in LLMs. It can also be used to enhance certain traits, personalities, moods, or preferences, such as emphasizing the concept of “brevity” or “reasoning” in any response an LLM generates. The team has made the method’s underlying code publicly available.

“LLMs clearly have a lot of these abstract concepts stored within them, in some representation,” Radhakrishnan says. “There are ways where, if we understand these representations well enough, we can build highly specialized LLMs that are still safe to use but really effective at certain tasks.”

This work was supported, in part, by the National Science Foundation, the Simons Foundation, the TILOS institute, and the U.S. Office of Naval Research. 

Originally from the small Balkan country of Montenegro, Strahinja (Strajo) Janjusevic says his life has unfolded in unexpected ways, for which he is deeply grateful. After graduating from high school, he was selected to represent his country in the United States, studying cyber operations and computer science at the U.S. Naval Academy in Annapolis, Maryland. He has since continued his cybersecurity studies and is currently a second-year master’s student in the Technology and Policy Program (TPP), hosted by the MIT Institute for Data, Systems, and Society (IDSS). His research with the MIT Laboratory for Information and Decision Systems (LIDS) and the MIT Maritime Consortium team aims to improve the cybersecurity of critical maritime infrastructure using artificial intelligence, considering both the technology and policy frameworks of solutions.

“My current research focuses on applying AI techniques to cybersecurity problems and examining the policy implications of these advancements, especially in the context of maritime cybersecurity,” says Janjusevic. “Representing my country at the highest levels of education and industry has given me a unique perspective on cybersecurity challenges.”

Janjusevic’s pathway from Montenegro to Maryland was created by a program that allows selected students from allied countries to attend the U.S. Naval Academy. Janjusevic graduated with a dual bachelor’s degree in cyber operations and computer science. His undergraduate experience provided opportunities to collaborate with the U.S. military and the National Security Agency, exposing him to high-level cybersecurity operations and fueling his interest in tackling complex cybersecurity challenges. During his undergraduate studies, he also interned with Microsoft, developing tools for cloud incident response, and with NASA, visualizing solar data for research applications.

Following his graduation, he realized that he still needed more knowledge, particularly in the area of AI and cybersecurity. TPP appealed to him immediately because of its dual emphasis on rigorous engineering innovation and the policy analysis needed to deploy it effectively. Janjusevic’s experiences at TPP have been a big change from his time at the U.S. Naval Academy, with a different pace and environment. He has especially appreciated being able to broaden his understanding about a variety of research domains and apply the discipline and knowledge he earned during his time at the academy.

“My TPP experience has been amazing,” says Janjusevic. “The cohort is really small, so it feels like a family, and everyone is working on diverse, high-impact problems.”

Mitigating the risks of emerging technologies

Janjusevic’s thesis brings together disciplines of cybersecurity, AI and deep learning, and control theory and physics, focusing on securing maritime cyber-physical systems — in particular, large legacy ships. The hacking of these ships’ networks can result in substantial damage to national security, as well as serious economic effects.

“Strajo is working to outsmart maritime GPS spoofing,” says Saurabh Amin, the Edmund K. Turner Professor in Civil Engineering. “Such attacks have already lured vessels off course in contested waters. His approach layers physics-based trajectory models with deep learning, catching threats that no single method can detect alone. His expertise has been very helpful in advancing our work on threat modeling and attack detection.”

The research utilizes advanced threat modeling and vessel dynamics to train AI systems to distinguish between legitimate maneuvers and spoofed signals. It involves building a framework that employs an internal LSTM (long short-term memory) autoencoder to analyze signal integrity, while simultaneously using a physics-based forecaster to predict the vessel’s movement based on environmental factors like wind and the sea state. By comparing these predictions against reported GPS positions, the system can effectively distinguish between natural sensor noise and malicious spoofing attacks. This hybrid framework is designed to empower, not replace, human operators, providing verified navigation data that allows watch standers to distinguish technical glitches from strategic cyberattacks.

Janjusevic has been able to enhance his academic research with industry experience. In summer 2025, he interned with the Network Detection team at the AI cybersecurity company Vectra AI. There, he investigated potential threats new technologies can bring, particularly AI agents and the model context protocol (MCP) — the emerging standard for AI agent communication. His research demonstrated how this technology could be repurposed for autonomous hacking operations and advanced command and control. This work on the security risks of agentic AI was recently presented in the preprint, “Hiding in the AI Traffic: Abusing MCP for LLM-Powered Agentic Red Teaming.”

“I was able to gain practical insights and hands-on experience into how a data science team uses AI models to detect anomalies in a network,” says Janjusevic. “This work within industry directly informed the anomaly detection models in my research.”

International policy perspective

“Strajo brings not just a high level of intelligence and energy to his work on cyber-physical security for merchant vessels, but also a strong instinct from his Navy training that resonates deeply with the research effort and grounds it in actionable policy,” says Fotini Christia, the Ford International Professor of the Social Sciences, director of IDSS, and a leader of the MIT Maritime Consortium.

Janjusevic participates in the cybersecurity efforts of the Maritime Consortium, a collaboration between academia, industry, and regulatory agencies focused on developing technological solutions, industry standards, and policies. The consortium includes cooperation with some international members, including from Singapore and South Korea.

“In AI cybersecurity, the policy element is really important, as the field is so fast-moving and the consequences of hacking can be so dangerous,” says Janjusevic. “I think there’s still a lot of need for policy work in this space.”

Janjusevic is also currently helping to organize two upcoming major conferences: the Harvard European Conference in February, which will convene officials and diplomats from across the globe, and the Technology and National Security Conference in April, a collaboration of Harvard and MIT that brings together top leaders from government, industry, and academia to tackle critical challenges in national security.

“I’m striving to find a position where I can influence and advance the cybersecurity field with AI, while at the same time leading collaboration and innovation between the United States and Montenegro,” says Janjusevic. “My goal is to be a bridge between Europe and the U.S. in this space of national security, AI, and cybersecurity, bringing my knowledge to both sides.”