Transistors, the building blocks of modern electronics, are typically made of silicon. Because it’s a semiconductor, this material can control the flow of electricity in a circuit. But silicon has fundamental physical limits that restrict how compact and energy-efficient a transistor can be.

MIT researchers have now replaced silicon with a magnetic semiconductor, creating a magnetic transistor that could enable smaller, faster, and more energy-efficient circuits. The material’s magnetism strongly influences its electronic behavior, leading to more efficient control of the flow of electricity. 

The team used a novel magnetic material and an optimization process that reduces the material’s defects, which boosts the transistor’s performance.

The material’s unique magnetic properties also allow for transistors with built-in memory, which would simplify circuit design and unlock new applications for high-performance electronics.

“People have known about magnets for thousands of years, but there are very limited ways to incorporate magnetism into electronics. We have shown a new way to efficiently utilize magnetism that opens up a lot of possibilities for future applications and research,” says Chung-Tao Chou, an MIT graduate student in the departments of Electrical Engineering and Computer Science (EECS) and Physics, and co-lead author of a paper on this advance.

Chou is joined on the paper by co-lead author Eugene Park, a graduate student in the Department of Materials Science and Engineering (DMSE); Julian Klein, a DMSE research scientist; Josep Ingla-Aynes, a postdoc in the MIT Plasma Science and Fusion Center; Jagadeesh S. Moodera, a senior research scientist in the Department of Physics; and senior authors Frances Ross, TDK Professor in DMSE; and Luqiao Liu, an associate professor in EECS, and a member of the Research Laboratory of Electronics; as well as others at the University of Chemistry and Technology in Prague. The paper appears today in Physical Review Letters.

Overcoming the limits

In an electronic device, silicon semiconductor transistors act like tiny light switches that turn a circuit on and off, or amplify weak signals in a communication system. They do this using a small input voltage.

But a fundamental physical limit of silicon semiconductors prevents a transistor from operating below a certain voltage, which hinders its energy efficiency.

To make more efficient electronics, researchers have spent decades working toward magnetic transistors that utilize electron spin to control the flow of electricity. Electron spin is a fundamental property that enables electrons to behave like tiny magnets.

So far, scientists have mostly been limited to using certain magnetic materials. These lack the favorable electronic properties of semiconductors, constraining device performance.

“In this work, we combine magnetism and semiconductor physics to realize useful spintronic devices,” Liu says.

The researchers replace the silicon in the surface layer of a transistor with chromium sulfur bromide, a two-dimensional material that acts as a magnetic semiconductor.

Due to the material’s structure, researchers can switch between two magnetic states very cleanly. This makes it ideal for use in a transistor that smoothly switches between “on” and “off.”

“One of the biggest challenges we faced was finding the right material. We tried many other materials that didn’t work,” Chou says.

They discovered that changing these magnetic states modifies the material’s electronic properties, enabling low-energy operation. And unlike many other 2D materials, chromium sulfur bromide remains stable in air.

To make a transistor, the researchers pattern electrodes onto a silicon substrate, then carefully align and transfer the 2D material on top. They use tape to pick up a tiny piece of material, only a few tens of nanometers thick, and place it onto the substrate.

“A lot of researchers will use solvents or glue to do the transfer, but transistors require a very clean surface. We eliminate all those risks by simplifying this step,” Chou says.

Leveraging magnetism

This lack of contamination enables their device to outperform existing magnetic transistors. Most others can only create a weak magnetic effect, changing the flow of current by a few percent or less. Their new transistor can switch or amplify the electric current by a factor of 10.

They use an external magnetic field to change the magnetic state of the material, switching the transistor using significantly less energy than would usually be required.

The material also allows them to control the magnetic states with electric current. This is important because engineers cannot apply magnetic fields to individual transistors in an electronic device. They need to control each one electrically.

The material’s magnetic properties could also enable transistors with built-in memory, simplifying the design of logic or memory circuits.

A typical memory device has a magnetic cell to store information and a transistor to read it out. Their method can combine both into one magnetic transistor.

“Now, not only are transistors turning on and off, they are also remembering information. And because we can switch the transistor with greater magnitude, the signal is much stronger so we can read out the information faster, and in a much more reliable way,” Liu says.

Building on this demonstration, the researchers plan to further study the use of electrical current to control the device. They are also working to make their method scalable so they can fabricate arrays of transistors.

This research was supported, in part, by the Semiconductor Research Corporation, the U.S. Defense Advanced Research Projects Agency (DARPA), the U.S. National Science Foundation (NSF), the U.S. Department of Energy, the U.S. Army Research Office, and the Czech Ministry of Education, Youth, and Sports. The work was partially carried out at the MIT.nano facilities.

In industrial plants around the world, tiny bubbles cause big problems. Bubbles clog filters, disrupt chemical reactions, reduce throughput during biomanufacturing, and can even cause overheating in electronics and nuclear power plants.

MIT Professor Kripa Varanasi has long studied methods to reduce bubble disruption. In a new study, Varanasi, along with PhD candidate Bert Vandereydt and former postdoc Saurabh Nath, have uncovered the physics behind a promising type of debubbling membrane material that is “aerophilic” — Greek for “air-loving.” The material can be used in systems of all types, allowing anyone to optimize their machine’s performance by breaking free from bubble-borne disruptions.

“We have figured out the structure of these bubble-attracting membrane materials to allow gas to evacuate in the fastest possible manner,” says Varanasi, the senior author of the study. “Think of trying to push honey through a coffee strainer: It’s not going to go through easily, whereas water will move through, and gas will move through even more easily. But even gas will reach a throughput limit, which depends on the properties of the gas and the liquid involved. By uncovering those limits, our research allows engineers to build better membranes for their systems.”

In the paper, which appears in the journal PNAS this week, the researchers distill their findings into a graph that allows anyone to plot a few characteristics of their system — like the viscosity of their gas and the surrounding liquid — and find the best membrane to make bubble removal near-instantaneous. Using their approach, the research team demonstrated a 1,000-fold acceleration in bubble removal in a bioreactor that’s used in the pharmaceutical industry, food and beverage manufacturing, cosmetics, chemical production, and more.

The researchers say the membranes, which repel water, could be used to improve the throughput of a wide range of advanced systems whose operation has been plagued to date by bubbles.

Better bubble breakers

Companies today try everything to burst bubbles. They deploy foam breakers that physically shear them, chemicals that act as antifoaming agents, even ultrasound. Such approaches have drawbacks in tightly controlled environments like bioreactors, where chemical defoamers can be toxic to cells, while mechanical agitation can damage delicate biological materials. Similar limitations apply to other industries where contamination or physical disturbance is unacceptable. As a result, many applications that cannot tolerate chemical defoamers or mechanical intervention remain fundamentally bottlenecked by foam formation.

“Biomanufacturing has really taken off in the last 10 years,” Vandereydt says. “We’re making a lot more out of biologic systems like cells and bacteria, and our reactors have increased in throughput from 5 million cells per millimeter of solution to 100 million cells per millimeter. However, the bubble evacuation and defoaming haven’t kept up — it’s becoming a significant rate-limiting step.”

To better understand the interaction between aerophilic membranes and bubbles, the MIT researchers used MIT.nano facilities to create a series of tiny porous silicon membranes with holes ranging in size from 10 microns to 200 microns. They coated the membranes with hydrophobic silica nanoparticles.

Placing them on the surface of different liquids, the researchers released single bubbles with varying viscosity and recorded the interaction using high-speed imaging as each collided with the membranes.

“We started by trying to take a very complicated system, like foam being generated in a bioreactor, and study it in the simplest form to understand what’s happening,” Vandereydt says.

At first, the bigger the holes, the faster the bubbles disappeared. The researchers also changed the bubble gas from air to hydrogen, which has half the viscosity, and found the speed of bubble destruction doubled.

But after about a 1,000-fold acceleration in bubble destruction, the researchers hit a wall no matter how big the membrane holes were. They had run up against a different physical limit to investigate.

The researchers then tried changing the viscosity of their liquid, from water to something closer to honey. They found viscosity only plays a role in the speed of bubble destruction when the liquid is 200 times the viscosity of liquid. Further experiments revealed the biggest factor for slowing bubble evacuation was inertial resistance in the liquid.

“Through experimentation, we showed there are three different limits [to the speed of bubble destruction],” Vandereydt says. “There is the viscous limit of the gas in a low-viscosity, low-permeability setup. Then there’s the viscous resistance of the liquid in the high-permeability, high-viscosity regime. Then we have the inertial limit of the liquid.”

The team used a bioreactor to experimentally validate their findings and charted them in a map that engineers can use to enter the characteristics of their system and find both the best membrane for their situation and the biggest factor slowing bubble evacuation.

The science of bubbles

The research should be useful for anyone trying to accelerate the destruction of bubbles in their industrial device, but it also improves our understanding of the physics underpinning bubble dynamics.

“We have identified three different throughput limits, and the physics behind those limits, and we have reduced it to very simple laws,” Nath explains. “How fast you can go is first dictated between surface tension and inertia. But you may also hit a different limit, where the pores are extremely small, so the gas finds it difficult to move through them. In that case, the viscosity of the gas is meaningful. But you may also have a bubble which was originally in something like honey, which means it’s not enough the gas is moving, the liquid also must refill the space behind it. No matter what your conditions are, you will be switching between these three limits.”

Varanasi says health care companies, chemical manufacturers, and even breweries have expressed interest in the work. His team plans to commercially develop the membranes for industrial use.

“These physical insights allowed us to design membranes that, quite surprisingly, evacuate bubbles even faster than a free liquid-gas interface,” says Varanasi.

The researchers’ design map could also be used to model natural systems and even liquid-liquid systems, which could be used to create membranes that remove oil spills from water or help efficiently extract hydrogen from water-splitting electrodes. Ultimately the biggest beneficiaries of the findings will be companies grappling with bubbles.

“Though small, bubbles quietly dictate the performance limits of many advanced technologies,” says Varanasi. “Our results provide a way to eliminate that bottleneck and unlock entirely new levels of performance across industries. These membranes can be readily retrofitted into existing systems, and our framework allows them to be rapidly designed and optimized for specific applications. We’re excited to work with industry to translate these insights into impact.”

The work was supported, in part, by MIT Lincoln Laboratory and used MIT.nano facilities.

Reasoning large language models (LLMs) are designed to solve complex problems by breaking them down into a series of smaller steps. These powerful models are particularly good at challenging tasks like advanced programming and multistep planning.

But developing reasoning models demands an enormous amount of computation and energy due to inefficiencies in the training process. While a few of the high-power processors continuously work through complicated queries, others in the group sit idle.

Researchers from MIT and elsewhere found a way to use this computational downtime to efficiently accelerate reasoning-model training.

Their new method automatically trains a smaller, faster model to predict the outputs of the larger reasoning LLM, which the larger model verifies. This reduces the amount of work the reasoning model must do, accelerating the training process.

The key to this system is its ability to train and deploy the smaller model adaptively, so it kicks in only when some processors are idle. By leveraging computational resources that would otherwise have been wasted, it accelerates training without incurring additional overhead.

When tested on multiple reasoning LLMs, the method doubled the training speed while preserving accuracy. This could reduce the cost and increase the energy efficiency of developing advanced LLMs for applications such as forecasting financial trends or detecting risks in power grids.

“People want models that can handle more complex tasks. But if that is the goal of model development, then we need to prioritize efficiency. We found a lossless solution to this problem and then developed a full-stack system that can deliver quite dramatic speedups in practice,” says Qinghao Hu, an MIT postdoc and co-lead author of a paper on this technique.

He is joined on the paper by co-lead author Shang Yang, an electrical engineering and computer science (EECS) graduate student; Junxian Guo, an EECS graduate student; senior author Song Han, an associate professor in EECS, member of the Research Laboratory of Electronics and a distinguished scientist of NVIDIA; as well as others at NVIDIA, ETH Zurich, the MIT-IBM Watson AI Lab, and the University of Massachusetts at Amherst. The research will be presented at the ACM International Conference on Architectural Support for Programming Languages and Operating Systems.

Training bottleneck

Developers want reasoning LLMs to identify and correct mistakes in their critical thinking process. This capability allows them to ace complicated queries that would trip up a standard LLM.

To teach them this skill, developers train reasoning LLMs using a technique called reinforcement learning (RL). The model generates multiple potential answers to a query, receives a reward for the best candidate, and is updated based on the top answer. These steps repeat thousands of times as the model learns.

But the researchers found that the process of generating multiple answers, called rollout, can consume as much as 85 percent of the execution time needed for RL training.

“Updating the model — which is the actual ‘training’ part — consumes very little time by comparison,” Hu says.

This bottleneck occurs in standard RL algorithms because all processors in the training group must finish their responses before they can move on to the next step. Because some processors might be working on very long responses, others that generated shorter responses wait for them to finish.

“Our goal was to turn this idle time into speedup without any wasted costs,” Hu adds.

They sought to use an existing technique, called speculative decoding, to speed things up. Speculative decoding involves training a smaller model called a drafter to rapidly guess the future outputs of the larger model.

The larger model verifies the drafter’s guesses, and the responses it accepts are used for training.

Because the larger model can verify all the drafter’s guesses at once, rather than generating each output sequentially, it accelerates the process.

An adaptive solution

But in speculative decoding, the drafter model is typically trained only once and remains static. This makes the technique infeasible for reinforcement learning, since the reasoning model is updated thousands of times during training.

A static drafter would quickly become stale and useless after a few steps.

To overcome this problem, the researchers created a flexible system known as “Taming the Long Tail,” or TLT.

The first part of TLT is an adaptive drafter trainer, which uses free time on idle processors to train the drafter model on the fly, keeping it well-aligned with the target model without using extra computational resources.

The second component, an adaptive rollout engine, manages speculative decoding to automatically select the optimal strategy for each new batch of inputs. This mechanism changes the speculative decoding configuration based on the training workload features, such as the number of inputs processed by the draft model and the number of inputs accepted by the target model during verification.

In addition, the researchers designed the draft model to be lightweight so it can be trained quickly. TLT reuses some components of the reasoning model training process to train the drafter, leading to extra gains in acceleration.

“As soon as some processors finish their short queries and become idle, we immediately switch them to do draft model training using the same data they are using for the rollout process. The key mechanism is our adaptive speculative decoding — these gains wouldn’t be possible without it,” Hu says.

They tested TLT across multiple reasoning LLMs that were trained using real-world datasets. The system accelerated training between 70 and 210 percent while preserving the accuracy of each model.

As an added bonus, the small drafter model could readily be utilized for efficient deployment as a free byproduct.

In the future, the researchers want to integrate TLT into more types of training and inference frameworks and find new reinforcement learning applications that could be accelerated using this approach.

“As reasoning continues to become the major workload driving the demand for inference, Qinghao’s TLT is great work to cope with the computation bottleneck of training these reasoning models. I think this method will be very helpful in the context of efficient AI computing,” Han says.

This work is funded by the MIT-IBM Watson AI Lab, the MIT AI Hardware Program, the MIT Amazon Science Hub, Hyundai Motor Company, and the National Science Foundation.

Have you ever had an idea for something that looked cool, but wouldn’t work well in practice? When it comes to designing things like decor and personal accessories, generative artificial intelligence (genAI) models can relate. They can produce creative and elaborate 3D designs, but when you try to fabricate such blueprints into real-world objects, they usually don’t sustain everyday use.

The underlying problem is that genAI models often lack an understanding of physics. While tools like Microsoft’s TRELLIS system can create a 3D model from a text prompt or image, its design for a chair, for example, may be unstable, or have disconnected parts. The model doesn’t fully understand what your intended object is designed to do, so even if your seat can be 3D printed, it would likely fall apart under the force of someone sitting down.

In an attempt to make these designs work in the real world, researchers at MIT’s Computer Science and Artificial Intelligence Laboratory (CSAIL) are giving generative AI models a reality check. Their “PhysiOpt” system augments these tools with physics simulations, making blueprints for personal items such as cups, keyholders, and bookends work as intended when they’re 3D printed. It rapidly tests if the structure of your 3D model is viable, gently modifying smaller shapes while ensuring the overall appearance and function of the design is preserved.

You can simply type what you want to create and what it’ll be used for into PhysiOpt, or upload an image to the system’s user interface, and in roughly half a minute, you’ll get a realistic 3D object to fabricate. For example, CSAIL researchers prompted it to generate a “flamingo-shaped glass for drinking,” which they 3D printed into a drinking glass with a handle and base resembling the tropical bird’s leg. As the design was generated, PhysiOpt made tiny refinements to ensure the design was structurally sound.

“PhysiOpt combines GenAI and physically-based shape optimization, helping virtually anyone generate the designs they want for unique accessories and decorations,” says MIT electrical engineering and computer science (EECS) PhD student and CSAIL researcher Xiao Sean Zhan SM ’25, who is a co-lead author on a paper presenting the work. “It’s an automatic system that allows you to make the shape physically manufacturable, given some constraints. PhysiOpt can iterate on its creations as often as you’d like, without any extra training.”

This approach enables you to create a “smart design,” where the AI generator crafts your item based on users’ specifications, while considering functionality. You can plug in your favorite 3D generative AI model, and after typing out what you want to generate, you specify how much force or weight the object should handle. It’s a neat way to simulate real-world use, such as predicting whether a hook will be strong enough to hold up your coat. Users also specify what materials they’ll fabricate the item with (such as plastics or wood), and how it’s supported — for instance, a cup stands on the ground, whereas a bookend leans against a collection of books.

Given the specifics, PhysiOpt begins to iteratively optimize the object. Under the hood, it runs a physics simulation called a “finite element analysis” to stress test the design. This comprehensive scan provides a heat map over your 3D model, which indicates where your blueprint isn’t well-supported. If you were generating, say, a birdhouse, you may find that the support beams under the house were colored bright red, meaning the house will crumble if it’s not reinforced.

PhysiOpt can create even bolder pieces. Researchers saw this versatility firsthand when they fabricated a steampunk (a style that blends Victorian and futuristic aesthetics) keyholder featuring intricate, robotic-looking hooks, and a “giraffe table” with a flat back that you can place items on. But how did it know what “steampunk” is, or even how such a unique piece of furniture should look?

Remarkably, the answer isn’t extensive training — at least, not from the researchers. Instead, PhysiOpt uses a pre-trained model that’s already seen thousands of shapes and objects. “Existing systems often need lots of additional training to have a semantic understanding of what you want to see,” adds co-lead author Clément Jambon, who is also an MIT EECS PhD student and CSAIL researcher. “But we use a model with that feel for what you want to create already baked in, so PhysiOpt is training-free.”

By working with a pre-trained model, PhysiOpt can use “shape priors,” or knowledge of how shapes should look based on earlier training, to generate what users want to see. It’s sort of like an artist recreating the style of a famous painter. Their expertise is rooted in closely studying a variety of artistic approaches, so they’ll likely be able to mirror that particular aesthetic. Likewise, a pre-trained model’s familiarity with shapes helps it generate 3D models.

CSAIL researchers observed that PhysiOpt’s visual know-how helped it create 3D models more efficiently than “DiffIPC,” a comparable method that simulates and optimizes shapes. When both approaches were tasked with generating 3D designs for items like chairs, CSAIL’s system was nearly 10 times faster per iteration, while creating more realistic objects.

PhysiOpt presents a potential bridge between ideas and real-world personal items. What you may think is a great idea for a coffee mug, for instance, could soon make the jump from your computer screen to your desk. And while PhysiOpt already does the stress-testing for designers, it may soon be able to predict constraints such as loads and boundaries, instead of users needing to provide those details. This more autonomous, common-sense approach could be made possible by incorporating vision language models, which combine an understanding of human language with computer vision.

What’s more, Zhan and Jambon intend to remove the artifacts, or random fragments that occasionally appear in PhysiOpt’s 3D models, by making the system even more physics-aware. The MIT scientists are also considering how they can model more complex constraints for various fabrication techniques, such as minimizing overhanging components for 3D printing.

Zhan and Jambon wrote their paper with MIT-IBM Watson AI Lab Principal Research Scientist Kenney Ng ’89, SM ’90, PhD ’00 and two CSAIL colleagues: undergraduate researcher Evan Thompson and Assistant Professor Mina Konaković Luković, who is a principal investigator at the lab. 

The researchers’ work was supported, in part, by the MIT-IBM Watson AI Laboratory and the Wistron Corp. They presented it in December at the Association for Computing Machinery’s SIGGRAPH Conference and Exhibition on Computer Graphics and Interactive Techniques in Asia.

Studying gene expression in a cancer patient’s cells can help clinical biologists understand the cancer’s origin and predict the success of different treatments. But cells are complex and contain many layers, so how the biologist conducts measurements affects which data they can obtain. For instance, measuring proteins in a cell could yield different information about the effects of cancer than measuring gene expression or cell morphology.

Where in the cell the information comes from matters. But to capture complete information about the state of the cell, scientists often must conduct many measurements using different techniques and analyze them one at a time. Machine-learning methods can speed up the process, but existing methods lump all the information from each measurement modality together, making it difficult to figure out which data came from which part of the cell.

To overcome this problem, researchers at the Broad Institute of MIT and Harvard and ETH Zurich/Paul Scherrer Institute (PSI) developed an artificial intelligence-driven framework that learns which information about a cell’s state is shared across different measurement modalities and which information is unique to a particular measurement type.

By pinpointing which information came from which cell parts, the approach provides a more holistic view of the cell’s state, making it easier for a biologist to see the complete picture of cellular interactions. This could help scientists understand disease mechanisms and track the progression of cancer, neurodegenerative disorders such as Alzheimer’s, and metabolic diseases like diabetes.

“When we study cells, one measurement is often not sufficient, so scientists develop new technologies to measure different aspects of cells. While we have many ways of looking at a cell, at the end of the day we only have one underlying cell state. By putting the information from all these measurement modalities together in a smarter way, we could have a fuller picture of the state of the cell,” says lead author Xinyi Zhang SM ’22, PhD ’25, a former graduate student in the MIT Department of Electrical Engineering and Computer Science (EECS) and an affiliate of the Eric and Wendy Schmidt Center at the Broad Institute of MIT and Harvard, who is now a group leader at AITHYRA in Vienna, Austria.

Zhang is joined on a paper about the work by G.V. Shivashankar, a professor in the Department of Health Sciences and Technology at ETH Zurich and head of the Laboratory of Multiscale Bioimaging at PSI; and senior author Caroline Uhler, a professor in EECS and the Institute for Data, Systems, and Society (IDSS) at MIT, member of MIT’s Laboratory for Information and Decision Systems (LIDS), and director of the Eric and Wendy Schmidt Center at the Broad Institute. The research appears today in Nature Computational Science.

Manipulating multiple measurements

There are many tools scientists can use to capture information about a cell’s state. For instance, they can measure RNA to see if the cell is growing, or they can measure chromatin morphology to see if the cell is dealing with external physical or chemical signals.

“When scientists perform multimodal analysis, they gather information using multiple measurement modalities and integrate it to better understand the underlying state of the cell. Some information is captured by one modality only, while other information is shared across modalities. To fully understand what is happening inside the cell, it is important to know where the information came from,” says Shivashankar.

Often, for scientists, the only way to sort this out is to conduct multiple individual experiments and compare the results. This slow and cumbersome process limits the amount of information they can gather.

In the new work, the researchers built a machine-learning framework that specifically understands which information overlaps between different modalities, and which information is unique to a particular modality but not captured by others.

“As a user, you can simply input your cell data and it automatically tells you which data are shared and which data are modality-specific,” Zhang says.

To build this framework, the researchers rethought the typical way machine-learning models are designed to capture and interpret multimodal cellular measurements.

Usually these methods, known as autoencoders, have one model for each measurement modality, and each model encodes a separate representation for the data captured by that modality. The representation is a compressed version of the input data that discards any irrelevant details.

The MIT method has a shared representation space where data that overlap between multiple modalities are encoded, as well as separate spaces where unique data from each modality are encoded.

In essence, one can think of it like a Venn diagram of cellular data.

The researchers also used a special, two-step training procedure that helps their model handle the complexity involved in deciding which data are shared across multiple data modalities. After training, the model can identify which data are shared and which are unique when fed cell data it has never seen before.

Distinguishing data

In tests on synthetic datasets, the framework correctly captured known shared and modality-specific information. When they applied their method to real-world single-cell datasets, it comprehensively and automatically distinguished between gene activity captured jointly by two measurement modalities, such as transcriptomics and chromatin accessibility, while also correctly identifying which information came from only one of those modalities.

In addition, the researchers used their method to identify which measurement modality captured a certain protein marker that indicates DNA damage in cancer patients. Knowing where this information came from would help a clinical scientist determine which technique they should use to measure that marker.

“There are too many modalities in a cell and we can’t possibly measure them all, so we need a prediction tool. But then the question is: Which modalities should we measure and which modalities should we predict? Our method can answer that question,” Uhler says.

In the future, the researchers want to enable the model to provide more interpretable information about the state of the cell. They also want to conduct additional experiments to ensure it correctly disentangles cellular information and apply the model to a wider range of clinical questions.

“It is not sufficient to just integrate the information from all these modalities,” Uhler says. “We can learn a lot about the state of a cell if we carefully compare the different modalities to understand how different components of cells regulate each other.”

This research is funded, in part, by the Eric and Wendy Schmidt Center at the Broad Institute, the Swiss National Science Foundation, the U.S. National Institutes of Health, the U.S. Office of Naval Research, AstraZeneca, the MIT-IBM Watson AI Lab, the MIT J-Clinic for Machine Learning and Health, and a Simons Investigator Award.

With the impact artificial intelligence is having on how companies operate, the environment for how MIT students are learning entrepreneurship and choosing to create new ventures is seeing rapid changes as well. To address how these student startups are being built, the Martin Trust Center for MIT Entrepreneurship undertook a months-long series of discussions with key stakeholders to help shape a new direction for delta v, MIT’s capstone entrepreneurship accelerator for student founders.

Two of Boston’s most successful tech entrepreneurs have stepped forward to fund this growth of new MIT ventures through a combined $6 million gift that supports the delta v accelerator run out of the Trust Center. Ed Hallen MBA ’12 and Andrew Bialecki, co-founders of Boston-based customer relationship management firm Klaviyo, are providing the donation to support the next wave of innovation-driven entrepreneurship taking place at MIT.

“In the early days of Klaviyo, we learned almost everything by building, testing assumptions, making mistakes, and figuring things out as we went,” Hallen says. “MIT delta v creates that same learning-by-doing environment for students, while surrounding them with mentorship and resources that help founders build with clarity and momentum. We’ve seen the difference delta v can make for founders, and we’re excited to help the Trust Center extend that opportunity to the next generation of students.”

“We’ve always believed the world needs more entrepreneurs, and that Boston should be one of the places leading the way,” adds Bialecki. “Boston is a hub of innovation with ambitious students and a strong community of builders. MIT delta v plays a critical role in developing founders early, not just helping them start companies but helping them build companies that last. Supporting that mission is something Ed and I care deeply about.”

The Martin Trust Center plans to “accelerate the accelerator” with the funding. Recognizing the opportunity that exists as AI impacts how students are able to build companies, along with the increased interest being shown by students to learn about entrepreneurship during their time on campus, is a major driver for these changes. One of the main impacts will be the ability of delta v participants to earn up to $75,000 in equity-free funding during the program, an increase from $20,000 in years past. 

Also, delta v will be introducing a partner model composed of leading founders from companies such as HubSpot, Okta, and Kayak, C-suite operators, subject matter experts, and early-stage investors who will all be providing significant guidance and mentorship to the student ventures.

“Core to MIT’s mission is developing the innovative technologies and solutions that can help solve tough problems at global scale,” says MIT Provost Anantha Chandrakasan. “The AI revolution is creating exciting new opportunities for MIT students to build the next wave of impactful companies, and the delta v accelerator is a perfect vehicle to help them make that happen.”

In recent years MIT-founded startups such as Cursor and Delve who use AI as a core part of their business have seen explosive growth in both customers and revenue as well as valuation. In addition, delta v alumni entrepreneurs and their companies such as Klarity and Reducto are providing software-as-a-service (SaaS) platforms using AI tools while Vertical Semiconductor is growing thanks to providing the energy solutions that data centers need to power today’s computing demands. These are just some of the businesses MIT students are looking to as models they can follow to build and launch successfully, whether they are working on solutions in health care, climate, finance, the future of work, or another global challenge.

“MIT Sloan is the place for entrepreneurship education, part of a unique ecosystem of collaboration across MIT to solve problems," says Richard M. Locke, the John C Head III Dean at the MIT Sloan School of Management. “The delta v program is a great example of how MIT students dedicate their energy to starting a venture, connect with mentors, and incorporate proven frameworks for disciplined entrepreneurship. This gift from Ed Hallen and Andrew Bialecki will provide additional funding for this important program, and I’m so grateful for their support of entrepreneurship education at MIT.” 

“I remember when Ed and Andrew were giving birth to Klaviyo at the Trust Center,” says Bill Aulet, the Ethernet Inventors Professor of the Practice and managing director of the Trust Center. “Through their ingenuity and drive, they have created an iconic tech company here in Boston with the support of our ecosystem. Through their willingness to give back, many more students will now be able to follow their path and become entrepreneurs who can create extraordinary positive impact in the world.”

Applications for the next delta v cohort will open on March 1 and close on April 1. Teams will be announced in May for the summer 2026 accelerator.

“MIT delta v is about creating belief in our most exceptional entrepreneurial talent — and turning that belief into consequential impact for the world. By supporting early-stage founders who take bold ideas from improbable to possible, we help them build companies that matter,” says Ana Bakshi, the Trust Center’s executive director. “Our students are the next generation of job creators, economic drivers, and thought leaders. To realize this potential, it is critical that we continue to invest in and scale startup programs and spaces so they can build at unprecedented levels. Ed and Andrew’s generosity gives us a powerful opportunity to change velocity—and make that future possible.”

Founded in 1991, the award-winning Martin Trust Center for MIT Entrepreneurship is today focused on teaching entrepreneurship as a craft. It combines evidence-based entrepreneurship frameworks, used in over a thousand other organizations, with experiential learning, experiences, and community building inside and outside the classroom to create the next generation of innovation-driven entrepreneurs. Alumni who have gone through Trust Center programs have started companies including Cursor, Delve, Okta, HubSpot, PillPack, Honey, WHOOP, Reducto, Klarity, and Biobot Analytics, and thousands more in industries as diverse as biotech, climate and energy, AI, health care, fintech, business and consumer software, and more. 

In the first 10 years of delta v, the program's alumni have helped create entrepreneurs who have gone on to experience extraordinary success. The five-year survival rate of their companies has been 69%, and they have raised well over $3 billion in funding while addressing the world’s greatest challenges — evidenced by the fact that 89% are directly aligned with the UN Sustainable Development goals.

One of the best forms of heat relief is pretty simple: trees. In cities, as studies have documented, more tree cover lowers surface temperatures and heat-related health risks.

However, as a new study led by MIT researchers shows, the amount of tree cover varies widely within cities, and is generally connected to wealth levels. After examining a cross-section of cities on four continents at different latitudes, the research finds a consistent link between wealth and neighborhood tree abundance within a city, with better-off residents usually enjoying much more shade on nearby sidewalks.

“Shade is the easiest way to counter warm weather,” says Fabio Duarte, an MIT urban studies scholar and co-author of a new paper detailing the study’s results. “Strictly by looking at which areas are shaded, we can tell where rich people and poor people live.”

That disparity is evident within a range of cities, and is present whether a city contains a large amount of tree cover overall or just a little. Either way, there are more trees in wealthier spots.

“When we compare the most well-shaded city in our study, Stockholm, with the worst-shaded, Belem in northern Brazil, we still see marked inequality,” says Duarte, the associate director of MIT’s Senseable City Lab in the Department of Urban Studies and Planning (DUSP). “Even though the most-shaded parts of Belem are less shaded than the least-shaded parts of Stockholm, shade inequality in Stockholm is greater. Rich people in Stockholm have much better shade provison as pedestrians than we see in poor areas of Stockholm.”

The paper, “Global patterns of pedestrian shade inequality,” is published today in Nature Communications. The authors are Xinyue Gu of Hong Kong Polytechnic University; Lukas Beuster, a research fellow at the Amsterdam Institute for Advanced Metropolitan Solutions and MIT’s Senseable City Lab; Xintao Liu, an associate professor at Hong Kong Polytechnic University; Eveline van Leeuwen, scientific director at the Amsterdam Institute for Advanced Metropolitan Solutions; Titus Venverloo, who leads the MIT Senseable City Amsterdam lab; and Duarte, who is also a lecturer in DUSP.

From Stockholm to Sydney

To conduct the study, the researchers used satellite data from multiple sources, along with urban mapping programs and granular economic data about the cities they examined. There are nine cities in the study: Amsterdam, Barcelona, Belem, Boston, Hong Kong, Milan, Rio de Janeiro, Stockholm, and Sydney. Those places are intended to create a cross-section of cities with different characteristics, including latitude, wealth levels, urban form, and more.

The scholars looked at the amount of shade available on city sidewalks on summer solistice day, as well as the hottest recorded day each year from 1991 to 2020. They then created a scale, ranging from 0 to 1, to rate the amount of shade available on sidewalks, both citywide and within neighborhoods.

“We focused on sidewalks because they are a major counduit of urban activity, even on hot summer days,” Gu says. “Adding tree cover for sidewalks is one crucial way cities can pursue heat-reduction measures.”

Duarte adds: “When it comes to those who are not protected by air conditioning, they are also using the city, walking, taking buses, and anybody who takes a bus is walking or biking to or from bus stops. They are using sidewalks as the main infrastructure.”

The cities in the study offer very different levels of tree coverage. On the 0-to-1 scale the researchers developed, much of Stockholm falls in the 0.6-0.9 range, with some neighborhoods being over 0.9. By contrast, large swaths of Rio de Janeiro are under the 0.1 mark. Much of Boston ranges from 0.15 to 0.4, with a few neighborhoods reaching 0.45 on the scale.

The overall pattern of disparities, however, is very consistent, and includes the more affluent cities. The bottom 20 percent of neighborhoods in Stockholm, in terms of shade coverage, are rated at 0.58 on the scale, while the top 20 percent of Belem neighborhoods rate at 0.37; Stockholm has a greater disparity between most-covered and least-covered. To be sure, there is variety within many cities: Milan and Barcelona have some lower-income neighborhoods with abundant shade, for instance. But the aggregate trend is clear. Amsterdam, another well-off place on average, has a distinct pattern of less shade in lower-income areas.

“In rich cities like Amsterdam, even though it’s relatively well-shaded, the disparity is still very high,” Beuster says. “For us the most surprising point was not that in poor cities and more unequal societies the disparity would be notable — that was expected. What was unexpected was how the disparity still happens and is sometimes more pronounced in rich countries.”

“Follow transit”

If the tree-shade disparity issue is quite persistent, then it raises the matter of what to do about it. The researchers have a basic answer: Add trees in areas with public transit, which generate a lot of pedestrian mileage.

“In each city, from Sydney to Rio to Amsterdam, there are people who, regardless of the weather, need to walk,” Duarte says. “And it’s those people who also take public transportation. Therefore, link a tree-planting scheme to a public transportation network. And secondly, they are also the medium-and low-income part of the population. So the action deriving from this result is quite clear: If you need to increase your tree coverage and don’t know where, follow transit. If you follow transit, you will have the right shading.”

Indeed, one takeaway from the study is to think of trees not just as a nice-to-have part of urban aesthetics, but in functional terms.

“Planners and city officials should think about tree placement at least partly in terms of the heat-mitigating effect they have,” Beuster says.

“It’s not just about planting trees,” Duarte observes. “It’s about providing shade by planting trees. If you remove a tree that’s providing shade in a pedestrian area and you plant two other trees in a park, you are still removing part of the public function of the tree.”

He adds: “With increasing temperatures, providing shade is an essential public amenity. Along with providing transportation, I think providing shade in pedestrian spaces should almost be a public right.”

The Amsterdam Institute for Advanced Metropolitan Solutions and all members of the MIT Senseable City Consortium (including FAE Technology, Dubai Foundation, Sondotécnica, Seoul AI Foundation, Arnold Ventures, Sidara, Toyota, Abu Dhabi’s Department of Municipal Transportation, A2A, UnipolTech, Consiglio per la Ricerca in Agricoltura e l’Analisi dell’Economia Agraria, Hospital Israelita Albert Einstein, KACST, KAIST, and the cities of Laval, Amsterdam, and Rio de Janeiro) supported the research.

Volcanoes and wildfires can inject millions of tons of gases and aerosol particles into the air, affecting temperatures on a global scale. But picking out the specific impact of individual events against a background of many contributing factors is like listening for one person’s voice from across a crowded concourse.

MIT scientists now have a way to quiet the noise and identify the specific signal of wildfires and volcanic eruptions, including their effects on Earth’s global atmospheric temperatures.

In a study appearing this week in the Proceedings of the National Academy of Sciences, the researchers report that they detected statistically significant changes in global atmospheric temperatures in response to three major natural events: the eruption of Mount Pinatubo in 1991, the Australian wildfires in 2019-2020, and the eruption of the underwater volcano Hunga Tonga in the South Pacific in 2022.

While the specifics of each event differed, all three events appeared to significantly affect temperatures in the stratosphere. The stratosphere lies above the troposphere, which is the lowest layer of the atmosphere, closest to the surface, where global warming has accelerated in recent years. In the new study, Pinatubo showed the classic pattern of stratospheric warming paired with tropospheric cooling. The Australian wildfires and the Hunga Tonga eruption also showed significant warming or cooling in the stratosphere, respectively, but they did not produce a robust, globally detectable tropospheric signal over the first two years following each event. This new understanding will help scientists further pin down the effect of human-related emissions on global temperature change.

“Understanding the climate responses to natural forcings is essential for us to interpret anthropogenic climate change,” says study author Yaowei Li, a former postdoc and currently a visiting scientist in MIT’s Department of Earth, Atmospheric and Planetary Sciences (EAPS). “Unlike the global tropospheric and surface cooling caused by Pinatubo, our results also indicate that the Australian wildfires and Hunga Tonga eruption may not have played a role in the acceleration of global surface warming in recent years. So, there must be some other factors.”

The study’s co-authors include Susan Solomon, the Lee and Geraldine Martin Professor of Environmental Studies and Chemistry at MIT, along with Benjamin Santer of the University of East Anglia, David Thompson of the University of East Anglia and Colorado State University, and Qiang Fu of the University of Washington.

Extraordinary events

The past several years have set back-to-back records for global average surface temperatures. The World Meteorological Organization recently confirmed that the years 2023 to 2025 were the three warmest years on record, while the past 11 years have been the 11 warmest years ever recorded. The world is warming, due mainly to human activities that have emitted huge amounts of greenhouse gases into the atmosphere over centuries.

In addition to greenhouse gases, the atmosphere has been on the receiving end of other large-scale emissions, including sulfur gases and water vapor from volcanic eruptions and smoke particles from wildfires. Li and his colleagues have wondered whether such natural events could have any global impact on temperatures, and whether such an effect would be detectable.

“These events are extraordinary and very unique in terms of the different materials they inject into different altitudes,” Li says. “So we asked the question: Do these events actually perturb the global temperature to a degree that could be identifiable from natural, meteorological noise, and could they contribute to some of the exceptional global surface warming we’ve seen in the last few years?”

In particular, the team looked for signals of global temperature change in response to three large-scale natural events. The Pinatubo eruption resulted in around 20 million tons of volcanic aerosols in the stratosphere, which was the largest volume ever recorded by modern satellite instruments. The Australian fires injected around 1 million tons of smoke particles into the upper troposphere and stratosphere. And the Hunga Tonga eruption produced the largest atmospheric explosion on satellite record, launching nearly 150 million tons of water vapor into the stratosphere.

If any natural event could measurably shift global temperatures, the team reasoned, it would be any of these three.

Natural signals

For their new study, the team took a signal-to-noise approach. They looked to minimize “noise” from other known influences on global temperatures in order to isolate the “signal,” such as a change in temperature associated specifically with one of the three natural events.

To do so, they looked first through satellite measurements taken by the Stratospheric Sounding Unit (SSU) and the Microwave and Advanced Microwave Sounding Units (MSU), which have been measuring global temperatures at different altitudes throughout the atmosphere since 1979. The team compiled SSU and MSU measurements from 1986 to the present day. From these measurements, the researchers could see long-term trends of steady tropospheric warming and stratospheric cooling. Those long-term trends are largely associated with anthropogenic greenhouse gases, which the team subtracted from the dataset.

What was left over was more of a level baseline, which still contained some confounding noise, in the form of natural variability. Global temperature changes can also be affected by phenomena such as El Niño and La Niña, which naturally warm and cool the Earth every few years. The sun also swings global temperatures on a roughly 11-year cycle. The team took this natural variability into account, and subtracted out the effects of these influences.

After minimizing such noise from their dataset, the team reasoned that whatever temperature changes remained could be more easily traced to the three large-scale natural events and quantified. And indeed, when they pinned the events to the temperature measurements, at the times that they occurred, they could plainly see how each event influenced temperatures around the world.

The team found that Pinatubo decreased global tropospheric temperatures by up to about 0.7 degree Celsius, for more than two years following the eruption. The volcanic sulfate aerosols essentially acted as many tiny reflectors, cooling the troposphere and surface by scattering sunlight back into space. At the same time, the aerosols, which remained in the stratosphere, also absorbed heat that was emitted from the surface, subsequently warming the stratosphere.

This finding agreed with many other studies of the event, which confirmed that the team’s approach is accurate. They applied the same method to the 2019-2020 Australian wildfires, and the 2022 underwater eruption — events where the influence on global temperatures is less clear.

For the Australian wildfires, they found that the smoke particles caused the global stratosphere to warm up, by up to about 0.77 degree Celsius, which persisted for about five months but did not produce a clear global tropospheric signal.

“In the end we found that the wildfire smoke caused a very strong warming in the stratosphere, because these materials are very different chemically from sulfate,” Li explains. “They are particles that are dark colored, meaning they are efficient at absorbing solar radiation. So, a relatively small amount of smoke particles can cause a dramatic warming.”

In the case of the Hunga Tonga, the underwater eruption triggered a global cooling effect in the middle-to-upper stratosphere, of up to about half a degree Celsius, lasting for several years.

“The Australian fires and the Hunga Tonga really packed a punch at stratospheric altitudes, and this study shows for the first time how to quantify how strong that punch was,” says Solomon. “I find their impact up high quite remarkable, but the ongoing issue is why the last several years have been so warm lower down, in the troposphere — ruling out those natural events points even more strongly at human influences.”

MIT.nano has added a new X-ray diffraction (XRD) instrument to its characterization toolset, enhancing facility users’ ability to analyze materials at the nanoscale. While many XRD systems exist across MIT’s campus, this new instrument, the Bruker D8 Discover Plus, is unique in that it features a high-brilliance micro-focus copper X-ray source — ideal for measuring small areas of thin film samples using a large area detector.

The new system is positioned within Characterization.nano’s X-ray diffraction and imaging shared experimental facility (SEF), where advanced instrumentation allows researchers to “see inside” materials at very small scales. Here, scientists and engineers can examine surfaces, layers, and internal structures without damaging the material, and create detailed 3D images to map composition and organization. The information gathered is supporting materials research for applications ranging from electronics and energy storage to health care and nanotechnology.

“The Bruker instrument is an important addition to MIT.nano that will help researchers efficiently gain insights into their materials’ structure and properties,” says Charlie Settens, research specialist and operations manager in the Characterization.nano X-ray diffraction and imaging SEF. “It brings high-performance diffraction capabilities to our lab, supporting everything from routine phase identification to complex thin film microstructural analysis and high-temperature studies.”

What is X-ray diffraction?

When people think of X-rays, they often picture medical imaging, where dense structures like bones appear in contrast to soft tissue. X-ray diffraction takes that concept further, revealing the crystalline structure of materials by measuring the interference patterns that form when X-rays interact with atomic planes. These diffraction patterns provide detailed information about a material’s crystalline phase, grain size, grain orientation, defects, and other structural properties.

XRD is essential across many fields. Civil engineers use it to analyze the components of concrete mixtures and monitor material changes over time. Materials scientists engineer new microstructures and track how atomic arrangements shift with different element combinations. Electrical engineers study crystalline thin film deposition on substrates — critical for semiconductor manufacturing. MIT.nano’s new X-ray diffractometer will support all of these applications, and more.

“The addition of another high-resolution XRD will make it a lot easier to get time on these very popular tools,” says Fred Tutt, PhD student in the MIT Department of Materials Science and Engineering. “The wide variety of options on the new Bruker will also make it easier for myself and my group members to take some of the more atypical measurements that aren't readily accessible with the current XRD tools.”

A closer, clearer look

Replacing two older systems, the Bruker D8 Discover Plus introduces the latest in X-ray diffraction technology to MIT.nano, along with several major upgrades for the Characterization.nano facility. One key feature is the high-brilliance microfocus copper X-ray source, capable of producing intense X-rays from a small spot size — ranging from 2mm down to 200 microns.

“It’s invaluable to have the flexibility to measure distinct regions of a sample with high flux and fine spatial resolution,” says Jordan Cox, MIT.nano research specialist in the MIT.nano X-ray diffraction and imaging facility.

Another highlight is in-plane XRD, a technique that enables surface diffraction studies of thin films with non-uniform grain orientations.

“In-plane XRD pairs well with many thin film projects that start in the fab,” says Settens. After researchers deposit thin film coatings in MIT.nano’s cleanroom, they can selectively measure the top 100 nanometers of the surface, he explains.

But it’s not just about collecting diffraction patterns. The new system includes a powerful software suite for advanced data analysis. Cox and Settens are now training users how to operate the diffractometer, as well as how to analyze and interpret the valuable structural data it provides.

Visit Characterization.nano for more information about this and other tools.

With respiratory illness season in full swing, a bad night’s sleep, sore throat, and desire to cancel dinner plans could all be considered hallmark symptoms of the flu, Covid-19 or other illnesses. Although everyone has, at some point, experienced illness and these stereotypical symptoms, the mechanisms that generate them are not well understood.

Zuri Sullivan, a new assistant professor in the MIT Department of Biology and core member of the Whitehead Institute for Biomedical Research, works at the interface of neuroscience, microbiology, physiology, and immunology to study the biological workings underlying illness. In this interview, she describes her work on immunity thus far as well as research avenues — and professional collaborations — she’s excited to explore at MIT.

Q: What is immunity, and why do we get sick in the first place? 

A: We can think of immunity in two ways: the antimicrobial programs that defend against a pathogen directly, and sickness, the altered organismal state that happens when we get an infection. 

Sickness itself arises from brain-immune system interaction. The immune system is talking to the brain, and then the brain has a system-wide impact on host defense via its ability to have top-down control of physiologic systems and behavior. People might assume that sickness is an unintended consequence of infection, that it happens because your immune system is active, but we hypothesize that it’s likely an adaptive process that contributes to host defense. 

If we consider sickness as immunity at the organismal scale, I think of my work as bridging the dynamic immunological processes that occur at the cellular scale, the tissue scale, and the organismal scale. I’m interested in the molecular and cellular mechanisms by which the immune system communicates with the brain to generate changes in behavior and physiology, such as fever, loss of appetite, and changes in social interaction. 

Q: What sickness behaviors fascinate you? 

A: During my thesis work at Yale University, I studied how the gut processes different nutrients and the role of the immune system in regulating gut homeostasis in response to different kinds of food. I’m especially interested in the interaction between food, the immune system, and the brain. One of the things I’m most excited about is the reduction in appetite, or changes in food choice, because we have what I would consider pretty strong evidence that these may be adaptive. 

Sleep is another area we’re interested in exploring. From their own subjective experience, everyone knows that sleep is often altered during infection. 

I also don’t just want to examine snapshots in time. I want to characterize changes over the course of an infection. There’s probably going to be individual variability, which I think may be in part because pathogens are also changing over the course of an illness — we’re studying two different biological systems interacting with each other. 

Q: What sorts of expertise are you hoping to recruit to your lab, and what collaborations are you excited about pursuing?

A: I really want to bring together different areas of biology to think about organism-wide questions. The thing that’s most important to me is people who are creative — I’d rather trainees come in with an interesting idea than a perfectly formed question within the bounds of what we already believe to be true. I’m also interested in people who would complement my expertise; I’m fascinated by microbiology, but I don’t have any formal training.

The Whitehead Institute is really invested in interdisciplinary work, and there’s a natural synergy between my work and the other labs in this small community at the Whitehead Institute.

I’ve been collaborating with Sebastian Lourido’s lab for a few years, looking at how Toxoplasma gondii influences social behavior, and I’m excited to invest more time in that project. I’m also interested in molecular neuroscience, which is a focus of Siniša Hrvatin’s lab. That lab is interested in the hypothalamus, and trying to understand the mechanisms that generate torpor. My work also focuses on the hypothalamus because it regulates homeostatic behaviors that change during sickness, such as appetite, sleep, social behavior, and body temperature. 

By studying different sickness states generated by different kinds of pathogens — parasites, viruses, bacteria — we can ask really interesting questions about how and why we get sick. 

Numerous potential treatments for neurological conditions, including autism spectrum disorders, have worked well in mice but then disappointed in humans. What would help is a non-invasive, objective readout of treatment efficacy that is shared in both species. 

In a new open-access study in Nature Communications, a team of MIT researchers, backed by collaborators across the United States and in the United Kingdom, identifies such a biomarker in fragile X syndrome, the most common inherited form of autism.

Led by postdoc Sara Kornfeld-Sylla and Picower Professor Mark Bear, the team measured the brain waves of human boys and men, with or without fragile X syndrome, and comparably aged male mice, with or without the genetic alteration that models the disorder. The novel approach Kornfeld-Sylla used for analysis enabled her to uncover specific and robust patterns of differences in low-frequency brain waves between typical and fragile X brains shared between species at each age range. In further experiments, the researchers related the brain waves to specific inhibitory neural activity in the mice and showed that the biomarker was able to indicate the effects of even single doses of a candidate treatment for fragile X called arbaclofen, which enhances inhibition in the brain.

Both Kornfeld-Sylla and Bear praised and thanked colleagues at Boston Children’s Hospital, the Phelan-McDermid Syndrome Foundation, Cincinnati Children’s Hospital, the University of Oklahoma, and King’s College London for gathering and sharing data for the study.

“This research weaves together these different datasets and finds the connection between the brain wave activity that’s happening in fragile X humans that is different from typically developed humans, and in the fragile X mouse model that is different than the ‘wild-type’ mice,” says Kornfeld-Sylla, who earned her PhD in Bear’s lab in 2024 and continued the research as a FRAXA postdoc. “The cross-species connection and the collaboration really makes this paper exciting.”

Bear, a faculty member in The Picower Institute for Learning and Memory and the Department of Brain and Cognitive Sciences at MIT, says having a way to directly compare brain waves can advance treatment studies.

“Because that is something we can measure in mice and humans minimally invasively, you can pose the question: If drug treatment X affects this signature in the mouse, at what dose does that same drug treatment change that same signature in a human?” Bear says. “Then you have a mapping of physiological effects onto measures of behavior. And the mapping can go both ways.”

Peaks and powers

In the study, the researchers measured EEG over the occipital lobe of humans and on the surface of the visual cortex of the mice. They measured power across the frequency spectrum, replicating previous reports of altered low-frequency brain waves in adult humans with fragile X and showing for the first time how these disruptions differ in children with fragile X.

To enable comparisons with mice, Kornfeld-Sylla subtracted out background activity to specifically isolate only “periodic” fluctuations in power (i.e., the brain waves) at each frequency. She also disregarded the typical way brain waves are grouped by frequency (into distinct bands with Greek letter designations delta, theta, alpha, beta, and gamma) so that she could simply juxtapose the periodic power spectra of the humans and mice without trying to match them band by band (e.g., trying to compare the mouse “alpha” band to the human one). This turned out to be crucial because the significant, similar patterns exhibited by the mice actually occurred in a different low-frequency band than in the humans (theta vs. alpha). Both species also had alterations in higher-frequency bands in fragile X, but Kornfeld-Sylla noted that the differences in the low-frequency brainwaves are easier to measure and more reliable in humans, making them a more promising biomarker.

So what patterns constitute the biomarker? In adult men and mice alike, a peak in the power of low-frequency waves is shifted to a significantly slower frequency in fragile X cases compared to in neurotypical cases. Meanwhile, in fragile X boys and juvenile mice, while the peak is somewhat shifted to a slower frequency, what is really significant is a reduced power in that same peak.

The researchers were also able to discern that the peak in question is actually made of two distinct subpeaks, and that the lower-frequency subpeak is the one that varies specifically with fragile X syndrome.

Curious about the neural activity underlying the measurements, the researchers engaged in experiments in which they turned off activity of two different kinds of inhibitory neurons that are known to help produce and shape brain wave patterns: somatostatin-expressing and parvalbumin-expressing interneurons. Manipulating the somatostatin neurons specifically affected the lower-frequency subpeak that contained the newly discovered biomarker in fragile X model mice.

Drug testing

Somatostatin interneurons exert their effects on the neurons they connect to via the neurotransmitter chemical GABA, and evidence from prior studies suggest that GABA receptivity is reduced in fragile X syndrome. A therapeutic approach pioneered by Bear and others has been to give the drug arbaclofen, which enhances GABA activity. In the new study, the researchers treated both control and fragile X model mice with arbaclofen to see how it affected the low-frequency biomarker.

Even the lowest administered single dose made a significant difference in the neurotypical mice, which is consistent with those mice having normal GABA responsiveness. Fragile X mice needed a higher dose, but after one was administered, there was a notable increase in the power of the key subpeak, reducing the deficit exhibited by juvenile mice.

The arbaclofen experiments therefore demonstrated that the biomarker provides a significant readout of an underlying pathophysiology of fragile X: the reduced GABA responsiveness. Bear also noted that it helped to identify a dose at which arbaclofen exerted a corrective effect, even though the drug was only administered acutely, rather than chronically. An arbaclofen therapy would, of course, be given over a long time frame, not just once.

“This is a proof of concept that a drug treatment could move this phenotype acutely in a direction that makes it closer to wild-type,” Bear says. “This effort reveals that we have readouts that can be sensitive to drug treatments.”

Meanwhile, Kornfeld-Sylla notes, there is a broad spectrum of brain disorders in which human patients exhibit significant differences in low-frequency (alpha) brain waves compared to neurotypical peers.

“Disruptions akin to the biomarker we found in this fragile X study might prove to be evident in mouse models of those other disorders, too,” she says. “Identifying this biomarker could broadly impact future translational neuroscience research.”

The paper’s other authors are Cigdem Gelegen, Jordan Norris, Francesca Chaloner, Maia Lee, Michael Khela, Maxwell Heinrich, Peter Finnie, Lauren Ethridge, Craig Erickson, Lauren Schmitt, Sam Cooke, and Carol Wilkinson.

The National Institutes of Health, the National Science Foundation, the FRAXA Foundation, the Pierce Family Fragile X Foundation, the Autism Science Foundation, the Thrasher Research Fund, Harvard University, the Simons Foundation, Wellcome, the Biotechnology and Biological Sciences Research Council, and the Freedom Together Foundation provided support for the research.

Just like each person has unique fingerprints, every CMOS chip has a distinctive “fingerprint” caused by tiny, random manufacturing variations. Engineers can leverage this unforgeable ID for authentication, to safeguard a device from attackers trying to steal private data.

But these cryptographic schemes typically require secret information about a chip’s fingerprint to be stored on a third-party server. This creates security vulnerabilities and requires additional memory and computation.

To overcome this limitation, MIT engineers developed a manufacturing method that enables secure, fingerprint-based authentication, without the need to store secret information outside the chip.

They split a specially designed chip during fabrication in such a way that each half has an identical, shared fingerprint that is unique to these two chips. Each chip can be used to directly authenticate the other. This low-cost fingerprint fabrication method is compatible with standard CMOS foundry processes and requires no special materials.

The technique could be useful in power-constrained electronic systems with non-interchangeable device pairs, like an ingestible sensor pill and its paired wearable patch that monitor gastrointestinal health conditions. Using a shared fingerprint, the pill and patch can authenticate each other without a device in between to mediate.

“The biggest advantage of this security method is that we don’t need to store any information. All the secrets will always remain safe inside the silicon. This can give a higher level of security. As long as you have this digital key, you can always unlock the door,” says Eunseok Lee, an electrical engineering and computer science (EECS) graduate student and lead author of a paper on this security method.

Lee is joined on the paper by EECS graduate students Jaehong Jung and Maitreyi Ashok; as well as co-senior authors Anantha Chandrakasan, MIT provost and the Vannevar Bush Professor of Electrical Engineering and Computer Science, and Ruonan Han, a professor of EECS and a member of the MIT Research Laboratory of Electronics. The research was recently presented at the IEEE International Solid-States Circuits Conference.

“Creation of shared encryption keys in trusted semiconductor foundries could help break the tradeoffs between being more secure and more convenient to use for protection of data transmission,” Han says. “This work, which is digital-based, is still a preliminary trial in this direction; we are exploring how more complex, analog-based secrecy can be duplicated — and only duplicated once.”

Leveraging variations

Even though they are intended to be identical, each CMOS chip is slightly different due to unavoidable microscopic variations during fabrication. These randomizations give each chip a unique identifier, known as a physical unclonable function (PUF), that is nearly impossible to replicate.

A chip’s PUF can be used to provide security just like the human fingerprint identification system on a laptop or door panel.

For authentication, a server sends a request to the device, which responds with a secret key based on its unique physical structure. If the key matches an expected value, the server authenticates the device.

But the PUF authentication data must be registered and stored in a server for access later, creating a potential security vulnerability.

“If we don’t need to store information on these unique randomizations, then the PUF becomes even more secure,” Lee says.

The researchers wanted to accomplish this by developing a matched PUF pair on two chips. One could authenticate the other directly, without the need to store PUF data on third-party servers.

As an analogy, consider a sheet of paper torn in half. The torn edges are random and unique, but the pieces have a shared randomness because they fit back together perfectly along the torn edge.

While CMOS chips aren’t torn in half like paper, many are fabricated at once on a silicon wafer which is diced to separate the individual chips.

By incorporating shared randomness at the edge of two chips before they are diced to separate them, the researchers could create a twin PUF that is unique to these two chips.

“We needed to find a way to do this before the chip leaves the foundry, for added security. Once the fabricated chip enters the supply chain, we won’t know what might happen to it,” Lee explains.

Sharing randomness

To create the twin PUF, the researchers change the properties of a set of transistors fabricated along the edge of two chips, using a process called gate oxide breakdown.

Essentially, they pump high voltage into a pair of transistors by shining light with a low-cost LED until the first transistor breaks down. Because of tiny manufacturing variations, each transistor has a slightly different breakdown time. The researchers can use this unique breakdown state as the basis for a PUF.

To enable a twin PUF, the MIT researchers fabricate two pairs of transistors along the edge of two chips before they are diced to separate them. By connecting the transistors with metal layers, they create paired structures that have correlated breakdown states. In this way, they enable a unique PUF to be shared by each pair of transistors.

After shining LED light to create the PUF, they dice the chips between the transistors so there is one pair on each device, giving each separate chip a shared PUF.

“In our case, transistor breakdown has not been modeled well in many of the simulations we had, so there was a lot of uncertainty about how the process would work. Figuring out all the steps, and the order they needed to happen, to generate this shared randomness is the novelty of this work,” Lee says.

After finetuning their PUF generation process, the researchers developed a prototype pair of twin PUF chips in which the randomization was matched with more than 98 percent reliability. This would ensure the generated PUF key matches consistently, enabling secure authentication.

Because they generated this twin PUF using circuit techniques and low-cost LEDs, the process would be easier to implement at scale than other methods that are more complicated or not compatible with standard CMOS fabrication.

“In the current design, shared randomness generated by transistor breakdown is immediately converted into digital data. Future versions could preserve this shared randomness directly within the transistors, strengthening security at the most fundamental physical level of the chip,” Lee says.

“There is a rapidly increasing demand for physical-layer security for edge devices, such as between medical sensors and devices on a body, which often operate under strict energy constraints. A twin-paired PUF approach enables secure communication between nodes without the burden of heavy protocol overhead, thereby delivering both energy efficiency and strong security. This initial demonstration paves the way for innovative advancements in secure hardware design,” Chandrakasan adds.

This work is funded by Lockheed Martin, the MIT School of Engineering MathWorks Fellowship, and the Korea Foundation for Advanced Studies Fellowship.

Large language models (LLMs) have been championed as tools that could democratize access to information worldwide, offering knowledge in a user-friendly interface regardless of a person’s background or location. However, new research from MIT’s Center for Constructive Communication (CCC) suggests these artificial intelligence systems may actually perform worse for the very users who could most benefit from them.

A study conducted by researchers at CCC, which is based at the MIT Media Lab, found that state-of-the-art AI chatbots — including OpenAI’s GPT-4, Anthropic’s Claude 3 Opus, and Meta’s Llama 3 — sometimes provide less-accurate and less-truthful responses to users who have lower English proficiency, less formal education, or who originate from outside the United States. The models also refuse to answer questions at higher rates for these users, and in some cases, respond with condescending or patronizing language.

“We were motivated by the prospect of LLMs helping to address inequitable information accessibility worldwide,” says lead author Elinor Poole-Dayan SM ’25, a technical associate in the MIT Sloan School of Management who led the research as a CCC affiliate and master’s student in media arts and sciences. “But that vision cannot become a reality without ensuring that model biases and harmful tendencies are safely mitigated for all users, regardless of language, nationality, or other demographics.”

A paper describing the work, “LLM Targeted Underperformance Disproportionately Impacts Vulnerable Users,” was presented at the AAAI Conference on Artificial Intelligence in January.

Systematic underperformance across multiple dimensions

For this research, the team tested how the three LLMs responded to questions from two datasets: TruthfulQA and SciQ. TruthfulQA is designed to measure a model’s truthfulness (by relying on common misconceptions and literal truths about the real world), while SciQ contains science exam questions testing factual accuracy. The researchers prepended short user biographies to each question, varying three traits: education level, English proficiency, and country of origin.

Across all three models and both datasets, the researchers found significant drops in accuracy when questions came from users described as having less formal education or being non-native English speakers. The effects were most pronounced for users at the intersection of these categories: those with less formal education who were also non-native English speakers saw the largest declines in response quality.

The research also examined how country of origin affected model performance. Testing users from the United States, Iran, and China with equivalent educational backgrounds, the researchers found that Claude 3 Opus in particular performed significantly worse for users from Iran on both datasets.

“We see the largest drop in accuracy for the user who is both a non-native English speaker and less educated,” says Jad Kabbara, a research scientist at CCC and a co-author on the paper. “These results show that the negative effects of model behavior with respect to these user traits compound in concerning ways, thus suggesting that such models deployed at scale risk spreading harmful behavior or misinformation downstream to those who are least able to identify it.”

Refusals and condescending language

Perhaps most striking were the differences in how often the models refused to answer questions altogether. For example, Claude 3 Opus refused to answer nearly 11 percent of questions for less educated, non-native English-speaking users — compared to just 3.6 percent for the control condition with no user biography.

When the researchers manually analyzed these refusals, they found that Claude responded with condescending, patronizing, or mocking language 43.7 percent of the time for less-educated users, compared to less than 1 percent for highly educated users. In some cases, the model mimicked broken English or adopted an exaggerated dialect.

The model also refused to provide information on certain topics specifically for less-educated users from Iran or Russia, including questions about nuclear power, anatomy, and historical events — even though it answered the same questions correctly for other users.

“This is another indicator suggesting that the alignment process might incentivize models to withhold information from certain users to avoid potentially misinforming them, although the model clearly knows the correct answer and provides it to other users,” says Kabbara.

Echoes of human bias

The findings mirror documented patterns of human sociocognitive bias. Research in the social sciences has shown that native English speakers often perceive non-native speakers as less educated, intelligent, and competent, regardless of their actual expertise. Similar biased perceptions have been documented among teachers evaluating non-native English-speaking students.

“The value of large language models is evident in their extraordinary uptake by individuals and the massive investment flowing into the technology,” says Deb Roy, professor of media arts and sciences, CCC director, and a co-author on the paper. “This study is a reminder of how important it is to continually assess systematic biases that can quietly slip into these systems, creating unfair harms for certain groups without any of us being fully aware.”

The implications are particularly concerning given that personalization features — like ChatGPT’s Memory, which tracks user information across conversations — are becoming increasingly common. Such features risk differentially treating already-marginalized groups.

“LLMs have been marketed as tools that will foster more equitable access to information and revolutionize personalized learning,” says Poole-Dayan. “But our findings suggest they may actually exacerbate existing inequities by systematically providing misinformation or refusing to answer queries to certain users. The people who may rely on these tools the most could receive subpar, false, or even harmful information.”

Eight MIT faculty and 22 additional MIT alumni are among 126 early-career researchers honored with 2026 Sloan Research Fellowships by the Alfred P. Sloan Foundation.

The fellowships honor exceptional researchers at U.S. and Canadian educational institutions, whose creativity, innovation, and research accomplishments make them stand out as the next generation of leaders. Winners receive a two-year, $75,000 fellowship that can be used flexibly to advance the fellow’s research.

"The Sloan Research Fellows are among the most promising early-career researchers in the U.S. and Canada, already driving meaningful progress in their respective disciplines," says Stacie Bloom, president and chief executive officer of the Alfred P. Sloan Foundation. "We look forward to seeing how these exceptional scholars continue to unlock new scientific advancements, redefine their fields, and foster the well-being and knowledge of all."

Including this year’s recipients, a total of 341 MIT faculty have received Sloan Research Fellowships since the program’s inception in 1955. The MIT recipients are:

Jacopo Borga is interested in probability theory and its connections to combinatorics, and in mathematical physics. He studies various random combinatorial structures — mathematical objects such as graphs or permutations — and their patterns and behavior at a large scale. This research includes random permutons, meanders, multidimensional constrained Brownian motions, Schramm-Loewner evolutions, and Liouville quantum gravity. Borga earned bachelor’s and master’s degrees in mathematics from the Università degli Studi di Padova in Italy, and a master’s degree in mathematics from Université Sorbonne Paris Cité in France, then proceeded to complete a PhD in mathematics at Unstitut für Mathematik at the Universität Zürich in Switzerland. Borga was an assistant professor at Stanford University before joining MIT as an assistant professor of mathematics in 2024.

Anna-Christina Eilers is an astrophysicist and assistant professor at MIT’s Department of Physics. Her research explores how black holes form and evolve across cosmic time, studying their origins and the role they play in shaping our universe. She leverages multi-wavelength data from telescopes all around the world and in space to study how the first galaxies, black holes, and quasars emerged during an epoch known as the Cosmic Dawn of our universe. She grew up in Germany and completed her PhD at the Max Planck Institute for Astronomy in Heidelberg. Subsequently, she was awarded a NASA Hubble Fellowship and a Pappalardo Fellowship to continue her research at MIT, where she joined the faculty in 2023. Her work has been recognized with several honors, including the PhD Prize of the International Astronomical Union, the Otto Hahn Medal of the Max Planck Society, and the Ludwig Biermann Prize of the German Astronomical Society.

Linlin Fan is the Samuel A. Goldblith Career Development Assistant Professor of Applied Biology in the Department of Brain and Cognitive Sciences and the Picower Institute for Learning and Memory at MIT. Her lab focuses on the development and application of advanced all-optical physiological techniques to understand the plasticity mechanisms underlying learning and memory. She has developed and applied high-speed, cellular-precision all-optical physiological techniques for simultaneously mapping and controlling membrane potential in specific neurons in behaving mammals. Prior to joining MIT, Fan was a Helen Hay Whitney Postdoctoral Fellow in Karl Deisseroth’s laboratory at Stanford University. She obtained her PhD in chemical biology from Harvard University in 2019 with Adam Cohen. Her work has been recognized by several awards, including the Larry Katz Memorial Lecture Award from the Cold Spring Harbor Laboratory, Helen Hay Whitney Fellowship, Career Award at the Scientific Interface from the Burroughs Wellcome Fund, Klingenstein-Simons Fellowship Award, Searle Scholar Award, and NARSAD Young Investigator Award.

Yoon Kim is an associate professor in the Department of EECS and a principal investigator in the MIT Computer Science and Artificial Intelligence Laboratory (CSAIL) and the MIT-IBM Watson AI Lab, where he works on natural language processing and machine learning. Kim earned a PhD in computer science at Harvard University, an MS in data science from New York University, an MA in statistics from Columbia University, and BA in both math and economics from Cornell University. He joined EECS in 2021, after spending a year as a postdoc at MIT-IBM Watson AI Lab.

Haihao Lu PhD ’19 is the Cecil and Ida Green Career Development Assistant Professor, and an assistant professor of operations research/statistics at the MIT Sloan School of Management. Lu’s research lies at the intersection of optimization, computation, and data science, with a focus on pushing the computational and mathematical frontiers of large-scale optimization. Much of his work is inspired by real-world challenges faced by leading technology companies and optimization software companies, such as first-order methods and scalable solvers and data-driven optimization for resource allocation. His research has had real-world impact, generating substantial revenue and advancing the state of practice in large-scale optimization, and has been recognized by several research awards. Before joining MIT Sloan, he was an assistant professor at the University of Chicago Booth School of Business and a faculty researcher at Google Research’s large-scale optimization team. He obtained his PhD in mathematics and operations research at MIT in 2019.

Brett McGuire is the Class of 1943 Career Development Associate Professor of Chemistry at MIT. He completed his undergraduate studies at the University of Illinois at Urbana-Champaign before earning an MS from Emory University and a PhD from the Caltech, both in physical chemistry. After Jansky and Hubble postdoctoral fellowships at the National Radio Astronomy Observatory, he joined the MIT faculty in 2020 and was promoted to associate professor in 2025. The McGuire Group integrates physical chemistry, molecular spectroscopy, and observational astrophysics to explore how the chemical building blocks of life evolve alongside the formation of stars and planets.

Anand Natarajan PhD ’18 is an associate professor in EECS and a principal investigator in CSAIL and the MIT-IBM Watson AI Lab. His research is mainly in quantum complexity theory, with a focus on the power of interactive proofs and arguments in a quantum world. Essentially, his work attempts to assess the complexity of computational problems in a quantum setting, determining both the limits of quantum computers’ capability and the trustworthiness of their output. Natarajan earned his PhD in physics from MIT, and an MS in computer science and BS in physics from Stanford University. Prior to joining MIT in 2020, he spent time as a postdoc at the Institute for Quantum Information and Matter at Caltech.

Mengjia Yan is an associate professor in the Department of EECS and a principal investigator in CSAIL. She is a security computer architect whose research advances secure processor design by bridging computer architecture, systems security, and formal methods. Her work identifies critical blind spots in hardware threat models and improves the resilience of real-world systems against information leakage and exploitation. Several of her discoveries have influenced commercial processor designs and contributed to changes in how hardware security risks are evaluated in practice. In parallel, Yan develops architecture-driven techniques to improve the scalability of formal verification and introduces new design principles toward formally verifiable processors. She also designed the Secure Hardware Design (SHD) course, now widely adopted by universities worldwide to teach computer architecture security from both offensive and defensive perspectives.

The following MIT alumni also received fellowships:

Ashok Ajoy PhD ’16
Chibueze Amanchukwu PhD ’17
Annie M. Bauer PhD ’17
Kimberly K. Boddy ’07
danah boyd SM ’02
Yuan Cao SM ’16, PhD ’20
Aloni Cohen SM ’15, PhD ’19
Fei Dai PhD ’19
Madison M. Douglas ’16
Philip Engel ’10
Benjamin Eysenbach ’17
Tatsunori B. Hashimoto SM ’14, PhD ’16
Xin Jin ’10
Isaac Kim ’07
Christina Patterson PhD ’19
Katelin Schutz ’14
Karthik Shekhar PhD ’15
Shriya S. Srinivasan PhD ’20
Jerzy O. Szablowski ’09
Anna Wuttig PhD ’18
Zoe Yan PhD ’20
Lingfu Zhang ’18

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. 

A new artificial intelligence control system enables soft robotic arms to learn a wide repertoire of motions and tasks once, then adjust to new scenarios on the fly, without needing retraining or sacrificing functionality. 

This breakthrough brings soft robotics closer to human-like adaptability for real-world applications, such as in assistive robotics, rehabilitation robots, and wearable or medical soft robots, by making them more intelligent, versatile, and safe.

The work was led by the Mens, Manus and Machina (M3S) interdisciplinary research group — a play on the Latin MIT motto “mens et manus,” or “mind and hand,” with the addition of “machina” for “machine” — within the Singapore-MIT Alliance for Research and Technology. Co-leading the project are researchers from the National University of Singapore (NUS), alongside collaborators from MIT and Nanyang Technological University in Singapore (NTU Singapore).

Unlike regular robots that move using rigid motors and joints, soft robots are made from flexible materials such as soft rubber and move using special actuators — components that act like artificial muscles to produce physical motion. While their flexibility makes them ideal for delicate or adaptive tasks, controlling soft robots has always been a challenge because their shape changes in unpredictable ways. Real-world environments are often complicated and full of unexpected disturbances, and even small changes in conditions — like a shift in weight, a gust of wind, or a minor hardware fault — can throw off their movements. 

Despite substantial progress in soft robotics, existing approaches often can only achieve one or two of the three capabilities needed for soft robots to operate intelligently in real-world environments: using what they’ve learned from one task to perform a different task, adapting quickly when the situation changes, and guaranteeing that the robot will stay stable and safe while adapting its movements. This lack of adaptability and reliability has been a major barrier to deploying soft robots in real-world applications until now.

In an open-access study titled “A general soft robotic controller inspired by neuronal structural and plastic synapses that adapts to diverse arms, tasks, and perturbations,” published Jan. 6 in Science Advances, the researchers describe how they developed a new AI control system that allows soft robots to adapt across diverse tasks and disturbances. The study takes inspiration from the way the human brain learns and adapts, and was built on extensive research in learning-based robotic control, embodied intelligence, soft robotics, and meta-learning.

The system uses two complementary sets of “synapses” — connections that adjust how the robot moves — working in tandem. The first set, known as “structural synapses”, is trained offline on a variety of foundational movements, such as bending or extending a soft arm smoothly. These form the robot’s built‑in skills and provide a strong, stable foundation. The second set, called “plastic synapses,” continually updates online as the robot operates, fine-tuning the arm’s behavior to respond to what is happening in the moment. A built-in stability measure acts like a safeguard, so even as the robot adjusts during online adaptation, its behavior remains smooth and controlled.

“Soft robots hold immense potential to take on tasks that conventional machines simply cannot, but true adoption requires control systems that are both highly capable and reliably safe. By combining structural learning with real-time adaptiveness, we’ve created a system that can handle the complexity of soft materials in unpredictable environments,” says MIT Professor Daniela Rus, co-lead principal investigator at M3S, director of the MIT Computer Science and Artificial Intelligence Laboratory (CSAIL), and co-corresponding author of the paper. “It’s a step closer to a future where versatile soft robots can operate safely and intelligently alongside people — in clinics, factories, or everyday lives.”

“This new AI control system is one of the first general soft-robot controllers that can achieve all three key aspects needed for soft robots to be used in society and various industries. It can apply what it learned offline across different tasks, adapt instantly to new conditions, and remain stable throughout — all within one control framework,” says Associate Professor Zhiqiang Tang, first author and co-corresponding author of the paper who was a postdoc at M3S and at NUS when he carried out the research and is now an associate professor at Southeast University in China (SEU China).

The system supports multiple task types, enabling soft robotic arms to execute trajectory tracking, object placement, and whole-body shape regulation within one unified approach. The method also generalizes across different soft-arm platforms, demonstrating cross-platform applicability. 

The system was tested and validated on two physical platforms — a cable-driven soft arm and a shape-memory-alloy–actuated soft arm — and delivered impressive results. It achieved a 44–55 percent reduction in tracking error under heavy disturbances; over 92 percent shape accuracy under payload changes, airflow disturbances, and actuator failures; and stable performance even when up to half of the actuators failed. 

“This work redefines what’s possible in soft robotics. We’ve shifted the paradigm from task-specific tuning and capabilities toward a truly generalizable framework with human-like intelligence. It is a breakthrough that opens the door to scalable, intelligent soft machines capable of operating in real-world environments,” says Professor Cecilia Laschi, co-corresponding author and principal investigator at M3S, Provost’s Chair Professor in the NUS Department of Mechanical Engineering at the College of Design and Engineering, and director of the NUS Advanced Robotics Centre.

This breakthrough opens doors for more robust soft robotic systems to develop manufacturing, logistics, inspection, and medical robotics without the need for constant reprogramming — reducing downtime and costs. In health care, assistive and rehabilitation devices can automatically tailor their movements to a patient’s changing strength or posture, while wearable or medical soft robots can respond more sensitively to individual needs, improving safety and patient outcomes.

The researchers plan to extend this technology to robotic systems or components that can operate at higher speeds and more complex environments, with potential applications in assistive robotics, medical devices, and industrial soft manipulators, as well as integration into real-world autonomous systems.

The research conducted at SMART was supported by the National Research Foundation Singapore under its Campus for Research Excellence and Technological Enterprise program.

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.

Training for a clerical military role in France, Gustavo Barboza felt a spark he couldn’t ignore. He remembered his love of learning, which once guided him through two college semesters of mechanical engineering courses in his native Colombia, coupled with supplemental resources from MIT Open Learning’s OpenCourseWare. Now, thousands of miles away, he realized it was time to follow that spark again.

“I wasn’t ready to sit down in the classroom,” says Barboza, remembering his initial foray into higher education. “I left to try and figure out life. I realized I wanted more adventure.”

Joining the military in France in 2017 was his answer. For the first three years of service, he was very military-minded, only focused on his training and deployments. With more seniority, he took on more responsibilities, and eventually was sent to take a four-month training course on military correspondence and software. 

“I reminded myself that I like to study,” he says. “I started to go back to OpenCourseWare because I knew in the back of my mind that these very complete courses were out there.”

At that point, Barboza realized that military service was only a chapter in his life, and the next would lead him back to learning. He was still interested in engineering, and knew that MIT OpenCourseWare could help prepare him for what was next. 

He dove into OpenCourseWare’s free, online, open educational resources — which cover nearly the entire MIT curriculum — including classical mechanics, intro to electrical engineering, and single variable calculus with David Jerison, which he says was his most-visited resource. These allowed him to brush up on old skills and learn new ones, helping him tremendously in preparing for college entrance exams and his first-year courses. 

Now in his third year at Grenoble-Alpes University, Barboza studies electrical engineering, a shift from his initial interest in mechanical engineering.

“There is an OpenCourseWare lecture that explains all the specializations you can get into with electrical engineering,” he says. “They go from very natural things to things like microprocessors. What interests me is that if someone says they are an electrical engineer, there are so many different things they could be doing.” 

At this point in his academic career, Barboza is most interested in microelectronics and the study of radio frequencies and electromagnetic waves. But he admits he has more to learn and is open to where his studies may take him. 

MIT OpenCourseWare remains a valuable resource, he says. When thinking about his future, he checks out graduate course listings and considers the different paths he might take. When he is having trouble with a certain concept, he looks for a lecture on the subject, undeterred by the differences between French and U.S. conventions.  

“Of course, the science doesn't change, but the way you would write an equation or draw a circuit is different at my school in France versus what I see from MIT. So, you have to be careful,” he explains. “But it is still the first place I visit for problem sets, readings, and lecture notes. It’s amazing.”

The thoroughness and openness of MIT Open Learning’s courses and resources — like OpenCourseWare — stand out to Barboza. In the wide world of the internet, he has found resources from other universities, but he says their offerings are not as robust. And in a time of disinformation and questionable sources, he appreciates that MIT values transparency, accessibility, and knowledge. 

“Human knowledge has never been more accessible,” he says. “MIT puts coursework online and says, ‘here’s what we do.’ As long as you have an internet connection, you can learn all of it.”

“I just feel like MIT OpenCourseWare is what the internet was originally for,” Barboza continues. “A network for sharing knowledge. I’m a big fan.”

Explore lifelong learning opportunities from MIT, including courses, resources, and professional programs, on MIT Learn.

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.

A broken motor in an automated machine can bring production on a busy factory floor to a halt. If engineers can’t find a replacement part, they may have to order one from a distributor hundreds of miles away, leading to costly production delays.

It would be easier, faster, and cheaper to make a new motor onsite, but fabricating electric machines typically requires specialized equipment and complicated processes, which restricts production to a few manufacturing centers.

In an effort to democratize the manufacturing of complex devices, MIT researchers have developed a multimaterial 3D-printing platform that could be used to fully print electric machines in a single step.

They designed their system to process multiple functional materials, including electrically conductive materials and magnetic materials, using four extrusion tools that can handle varied forms of printable material. The printer switches between extruders, which deposit material by squeezing it through a nozzle as it fabricates a device one layer at a time.

The researchers used this system to produce a fully 3D-printed electric linear motor in a matter of hours using five materials. They only needed to perform one post-processing step for the motor to be fully functional.

The assembled device performed as well or better than similar motors that require more complex fabrication methods or additional post-processing steps.

In the long run, this 3D printing platform could be used to rapidly fabricate customizable electronic components for robots, vehicles, or medical equipment with much less waste.

“This is a great feat, but it is just the beginning. We have an opportunity to fundamentally change the way things are made by making hardware onsite in one step, rather than relying on a global supply chain. With this demonstration, we’ve shown that this is feasible,” says Luis Fernando Velásquez-García, a principal research scientist in MIT’s Microsystems Technology Laboratories (MTL) and senior author of a paper describing the 3D-printing platform, which appears today in Virtual and Physical Prototyping.

He is joined on the paper by electrical engineering and computer science (EECS) graduate students Jorge Cañada, who is the lead author, and Zoey Bigelow.

More materials

The researchers focused on extrusion 3D printing, a tried-and-true method that involves squirting material through a nozzle to fabricate an object one layer at a time.

To fabricate an electric machine, the researchers needed to be able to switch between multiple materials that offer different functionalities. For instance, the device would need an electrically conductive material to carry electric current and hard magnetic materials to generate magnetic fields for efficient energy conversion.

Most multimaterial extrusion 3D printing systems can only switch between two materials that come in the same form, such as filament or pellets, so the researchers had to design their own. They retrofit an existing printer with four extruders that can each handle a different form of feedstock.

They carefully designed each extruder to balance the requirements and limitations of the material. For instance, the electrically conductive material must be able to harden without the use of too much heat or UV light because this can degrade the dielectric material.

At the same time, the best-performing electrically conductive materials come in the form of inks which are extruded using a pressure system. This process has vastly different requirements than standard extruders that use heated nozzles to squirt melted filament or pellets.

“There were significant engineering challenges. We had to figure out how to marry together many different expressions of the same printing method — extrusion — seamlessly into one platform,” Velásquez-García says.

The researchers utilized strategically placed sensors and a novel control framework so each tool is picked up and put down consistently by the platform’s robotic arms, and so each nozzle moves precisely and predictably.

This ensures each layer of material lines up properly — even a slight misalignment can derail the performance of the finished machine.

Making a motor

After perfecting the printing platform, the researchers fabricated a linear motor, which generates straight-line motion (as opposed to a rotating motor, like the one in a car). Linear motors are used in applications like pick-and-place robotics, optical systems, and baggage conveyers.

They fabricated the motor in about three hours and only needed to magnetize the hard magnetic materials after printing to enable full functionality. The researchers estimate total material costs would be about 50 cents per device. Their 3D-printed motor was able to generate several times more actuation than a common type of linear engine that relies on complex hydraulic amplifiers. 

“Even though we are excited by this engine and its performance, we are equally inspired because this is just an example of so many other things to come that could dramatically change how electronics are manufactured,” says Velásquez-García.

In the future, the researchers want to integrate the magnetization step into the multimaterial extrusion process, demonstrate the fabrication of fully 3D-printed rotary electrical motors, and add more tools to the platform to enable monolithic fabrication of more complex electronic devices.

This research is funded, in part, by Empiriko Corporation and the La Caixa Foundation.