Les Perelman, expert in writing assessment and champion of writing education, dies at 77

Leslie “Les” Perelman, an influential figure in college writing assessment; a champion of writing instruction across all subject matters for over three decades at MIT; and a former MIT associate dean for undergraduate education, died on Nov. 12, 2025, at home in Lexington, Massachusetts. He was 77.

A Los Angeles native, Perelman attended the University of California at Berkeley, joining in its lively activist years, and in 1980 received his PhD in English from the University of Massachusetts at Amherst. After stints at the University of Southern California and Tulane University, he returned to Massachusetts — to MIT — in 1987, and stayed for the next 35 years.

Perelman became best known for his dogged critique of autograding systems and writing assessments that didn’t assess actual college writing. The Boston Globe dubbed him “The man who killed the SAT essay.” He told NPR that colleges “spend the first year deprogramming [students] from the five-paragraph essay.” 

His widow, MIT Professor Emerita Elizabeth Garrels, says that while attending a conference, Perelman — who was practically blind without his glasses — arranged to stand at one end of a room in order to “grade” essays held up for him on the other side. “He would call out the grade that each essay would likely receive on standardized scoring,” Garrels says. “And he was consistently right.” Perelman was doing what automatic scorers were: He was, he said in the NPR interview, “mirroring how automated or formulaic grading systems often reward form over substance.” 

Perelman also “ruffled a lot of feathers” in industry, says Garrels, with his 2020 paper documenting his BABEL (“Basic Automatic B.S. Essay Language”) Generator, which output nonsense that commercial autograders nevertheless gave top marks. He saved some of his most systematic criticism for autograders’ defenders in academia, at one point calling out peers at the University of Akron for the methodology in their widely-touted paper claiming autograders performed just as well as human graders. 

At least one service, though, E.T.S., partly welcomed Perelman’s critique by making its autograder available to him for testing. (Others, like Pearson and Vantage Learning, declined.) He discovered he could ace the tests, even when his essay included non-factual gibberish and typographical errors:

Teaching assistants are paid an excessive amount of money. The average teaching assistant makes six times as much money as college presidents. In addition, they often receive a plethora of extra benefits such as private jets, vacations in the south seas, a staring roles in motion pictures. Moreover, in the Dickens novel Great Expectation, Pip makes his fortune by being a teaching assistant. It doesn’t matter what the subject is, since there are three parts to everything you can think of.

MIT career

Within MIT, Perelman’s legacy was his push to embed writing instruction into the whole of MIT’s curriculum, not as standalone expository writing subjects, let alone as merely a writing exam that incoming students could use to pass out of writing subjects altogether. Supported by a $325,000 National Science Foundation grant, he convinced MIT to hire writing instructors who were also subject matter experts, often with STEM PhDs. They were tasked with collaborating with departments to plant writing instruction into both existing curricula and new subjects. That effort eventually became the Writing Across the Curriculum program (today named Writing, Rhetoric, and Professional Communication) with a staff of more than 30 instructors.

Building out the infrastructure wasn’t quick, however. Perelman’s successor, Suzanne Lane ’85, says it took him almost 15 years. It started with proving to others just how uneven writing instruction at MIT actually was. “A whole cohort of students who took a lot of writing classes or got communication instruction in various places would make great progress,” Lane says. “But it was definitely possible to get through all of MIT without doing much writing at all.” 

To bolster his case, Perelman turned to alumni surveys. “The surveys asked how well MIT prepared you for your career,” says Lane. “The technical skills scored really high, but — what is horribly termed, sometimes, as ‘soft skills’ — communication skills, collaboration, etc., these scored really high on importance to career, but really low on how well MIT had prepared them.”

In other words, MIT alumni knew their stuff but were bad at communicating it, at a cost to their careers.

This led Perelman and others to push for a new undergraduate communication requirement. That NSF grant supported a 1997 pilot, designing experiments for courses that would be communication-intensive. It was a huge success. Every department participated. It involved 24 subjects and roughly 300 students. MIT faculty, following “lively” discussion at an April 1999 faculty meeting, approved the proposal of the creation of a report on the communication requirement’s implementation, followed a year later by its formal passage, effective fall 2001.

From that initial pilot of 24, there are now nearly 300 subjects that count toward the requirement, from ​​class 1.013 (Senior Civil and Environmental Engineering Design) to 24.918 (Workshop in Linguistic Research).

Connections beyond MIT

Early in his career, Perelman worked with Vincent DiMarco, a literature scholar at the University of Massachusetts at Amherst, to publish “The Middle English Letter of Alexander to Aristotle” (Brill, 1978). With Wang Computers as publisher, he was a technical writer and project leader on the “DOS Release 3.30 User’s Reference Guide.” He edited a book and chapter on writing studies and assessment with New Jersey Institute of Technology professor Norbert Elliot. And in a project he was particularly proud of, he worked with the New South Wales Teachers Federation in 2018 to convince Australia to reject the adoption of an automated essay grading regime. 

“Les was brilliant, with a Talmudic way of asking questions and entering academic debates,” says Nancy Sommers, whose work on undergraduate writing assessment at Harvard University paralleled Perelman’s. “I loved the way his eyes sparkled when he was ready to rip an adversary or a colleague who wasn’t up to his quick mind and vast, encyclopedic knowledge.” 

Openness to rhetorical combat didn’t keep Perelman from being a wonderful friend, Sommers says, saying he once waited for her at the airline gate with a sandwich and a smile after a canceled flight. “That was Les, so gracious, generous, anticipating the needs of friends, always there to offer sustenance and friendship.”

Donations in Perelman’s name can be made to UNICEF’s work supporting children in Ukraine, the Lexington Refugee Assistance Program, Doctors Without Borders, and the Ash Grove Movie Finishing Fund.

de MIT News https://ift.tt/rgFz6qN

Coping with catastrophe

Each April in Japan, people participate in a tradition called “hanami,” or cherry-blossom viewing, where they picnic under the blooming trees. The tradition has a second purpose: The presence of people at these gatherings, often by water, helps solidify riverbanks and protect them from spring floods. The celebration has a dual purpose, by addressing, however incrementally, the threat of natural disaster.

The practice of creating things that also protect against disasters can be seen all over Japan, where many new or renovated school buildings have design features unfamiliar to students elsewhere. In Tokyo, one elementary school has a roof swimming pool that stores water and is used to help the building’s toilets flush, plus an additional rainwater catchment tank and exterior stairs leading to a large balcony that wraps around one side of the building.

Why? Well, Japan is prone to natural disasters, such as tsunamis, earthquakes, and flooding. The country’s schools often double as evacuation sites for local residents, and design practices increasingly reflect this. In normal times, the roof pool is where students learn to swim and helps keep the school cool, and the large balcony is used by spectators watching the adjacent school athletics field. In emergencies, water storage is crucial and exterior stairs help people ascend quickly to the gymnasium, built on the second floor — to keep evacuees safer during flooding.

Meanwhile, in one Tokyo district, rooftop solar power is now common. Some schools feature skylights and courtyards to bring in natural light. Again, these architectural features serve dual purposes. Solar power, for one, lowers annual operating costs, and it provides electricity even in case of grid troubles.

These are examples of what MIT scholar Miho Mazereeuw has termed “anticipatory design,” in which structures and spaces are built with dual uses, for daily living and for when crisis strikes.

“The idea is to have these proactive measures in place rather than being reactionary and jumping into action only after something has happened,” says Mazereeuw, an associate professor in MIT’s Department of Architecture and a leading expert on resilient design.

Now Mazereeuw has a new book on the subject, “Design Before Disaster: Japan’s Culture of Preparedness,” published by the University of Virginia Press. Based on many years of research, with extensive illustrations, Mazereeuw examines scores of successful design examples from Japan, both in terms of architectural features and the civic process that created them.

“I’m hoping there can be a culture shift,” Mazereeuw says. “Wherever you can invent design outcomes to help society be more resilient beforehand, it is not at exorbitant cost. You can design for exceptional everyday spaces but embed other infrastructure and flexibility in there, so when there is a flood event or earthquake, those buildings have more capability.”

Bosai and barbecue

Mazereeuw, who is also the head of MIT’s Urban Risk Lab, has been studying disaster preparedness for over 30 years. As part of the Climate Project at MIT, she is also one of the mission directors and has worked with communities around the world on resiliency planning.

Japan has a particularly well-established culture of preparedness, often referred to through the Japanese word “bosai.” Mazereeuw has been studying the country’s practices carefully since the 1990s. In researching the book, she has visited hundreds of sites in the country and talked to many officials, designers, and citizens along the way.

Indeed, Mazereeuw emphasizes, “A major theme in the book is connecting the top-down and bottom-up.” Some good design ideas come from planners and architects. Other have come from community groups and local residents. All these sources are important.

“The Japanese government does invest a lot in disaster research and recovery,” Mazereeuw says. “But I would hate for people in other countries to think this isn’t possible elsewhere. It’s the opposite. There are a lot of examples in here that don’t cost extra, because of careful design through community participation.”

As one example, Mazereeuw devotes a chapter of the book to public parks, which are often primary evacuation spaces for residents in case of emergency. Some have outdoor cooking facilities, which in normal times are used for, say, a weekend barbecue or local community events but are also there in case of emergency. Some parks also have water storage, or restroom facilities designed to expand if needed, and many serve as flood reservoirs, protecting the surrounding neighborhood.

“The barbecue facilities are a great example of dual use, connecting the everyday with disaster preparedness,” Mazereeuw says. “You can bring food into this beautiful park, so you’re used to using this space for cooking already. The idea is that your cognitive map of where you should go is connected to fun things you have done in the past.”

Some of the parks Mazereeuw surveys in the book are tiny pocket parks, which are also filled with useful resilience tools.

“Anticipatory design does not have to be monumental,” Mazereeuw writes in the book.

Negotiating through design

To be sure, some disaster mitigation measures are difficult to enact. In the Naiwan district of Kesennuma, as Mazereeuw outlines in the book, much of the local port area was destroyed in the 2011 tsunami, and the government wanted to build a seawall as part of the reconstruction plan. Some local residents and fishermen were unenthusiastic; a seawall could limit ocean access. Finally, after extended negotiations, designers created a seawall integrated into a new commercial district with cafes and stores, as well as new areas of public water access.

“This project used the power of design to negotiate between prefectural and local regulations, structural integrity and aesthetics, ocean access and safety,” Mazereeuw says.

Ultimately, working to build a coalition in support of resilience measures can help create more interesting and useful designs.

Other scholars have praised “Design Before Disaster.” Daniel P. Aldrich, a professor at Northeastern University, has called the book a “well-researched, clearly written investigation” into Japanese disaster-management practices, adding that any officials or citizens around the world “who seek to keep residents and communities safe from shocks of all kinds will learn something important from this book. It sets a high bar for future scholarship in the field.”

For her part, Mazereeuw emphasizes, “We can learn from the Japanese example, but it’s not a copy-paste thing. The book is so people can understand the essence of it and then create their own disaster preparedness culture and approach. This should be an all-hands process. Emergency management is not about relying on managers. It’s figuring out how we all play a part.”

de MIT News https://ift.tt/dCqAfRX

Featured video: Coding for underwater robotics

During a summer internship at MIT Lincoln Laboratory, Ivy Mahncke, an undergraduate student of robotics engineering at Olin College of Engineering, took a hands-on approach to testing algorithms for underwater navigation. She first discovered her love for working with underwater robotics as an intern at the Woods Hole Oceanographic Institution in 2024. Drawn by the chance to tackle new problems and cutting-edge algorithm development, Mahncke began an internship with Lincoln Laboratory's Advanced Undersea Systems and Technology Group in 2025. 

Mahncke spent the summer developing and troubleshooting an algorithm that would help a human diver and robotic vehicle collaboratively navigate underwater. The lack of traditional localization aids — such as the Global Positioning System, or GPS — in an underwater environment posed challenges for navigation that Mahncke and her mentors sought to overcome. Her work in the laboratory culminated in field tests of the algorithm on an operational underwater vehicle. Accompanying group staff to field test sites in the Atlantic Ocean, Charles River, and Lake Superior, Mahncke had the opportunity see her software in action in the real world.

"One of the lead engineers on the project had split off to go do other work. And she said, 'Here's my laptop. Here are the things that you need to do. I trust you to go do them.' And so I got to be out on the water as not just an extra pair of hands, but as one of the lead field testers," Mahncke says. "I really felt that my supervisors saw me as the future generation of engineers, either at Lincoln Lab or just in the broader industry."

Says Madeline Miller, Mahncke's internship supervisor: "Ivy's internship coincided with a rigorous series of field tests at the end of an ambitious program. We figuratively threw her right in the water, and she not only floated, but played an integral part in our program's ability to hit several reach goals."

Lincoln Laboratory's summer research program runs from mid-May to August. Applications are now open. 

Video by Tim Briggs/MIT Lincoln Laboratory | 2 minutes, 59 seconds

de MIT News https://ift.tt/qa1DYpx

Turning curiosity about engineering into careers

It’s not every day that aspiring teenage engineers can see firsthand how planes are built. But a collaboration between nonprofit Engineering Tomorrow, aerospace firm Boeing, and alumni of the MIT Leaders for Global Operations (LGO) program working at Boeing is aiming to turn curiosity about aerospace engineering into possible careers for young students.

Boeing is LGO’s longest-standing industry collaborator, hosting LGO internships, recruiting LGO alumni, and hosting plant treks for future engineers. Engineering Tomorrow, a nonprofit dedicated to inspiring the next generation of engineers, frames the U.S. engineering workforce shortage as an economic and national security issue — and says the shortage isn’t in just engineers with degrees, but also in trained operators and technicians. They also recognize that many kids often start as natural tinkerers, but get scared off by higher-level math.

To bring more kids into the engineering fold, the organization delivers no-cost engineering labs to middle and high school students by collaborating with influential mentors, such as LGO graduates at organizations like Boeing.

“We want to inspire students by exposing them to professional engineers to illustrate the pathways for them to be problem-solvers in society,” explains Alex Dickson, Engineering Tomorrow’s program coordinator. “The demand for engineers has just gone up dramatically. It’s about being competitive on a global scale. We try to illustrate to students that there are many pathways into these careers.”

How MIT LGO makes engineering dreams a reality

Engineering Tomorrow’s collaboration with MIT LGO grew organically, through a robust alumni network. One of the nonprofit’s board members, LGO alumna Kristine Budill SM ’93, recognized a shared interest: the sizable Boeing LGO community wanted concrete ways to connect more directly with communities, and Engineering Tomorrow does just that.

Budill connected the organization with fellow LGO alumnus Cameron Hoffman MBA ’24, SM ’24, a Boeing manufacturing strategy manager who helped translate that shared mission into a real-world opportunity: an on-site Boeing experience that made engineering tangible for high school students.

The result: One lucky high school engineering design class from Mercer Island, Washington, recently got to experience Boeing 737s being built in person. In November 2025, 30 ninth graders at Mercer Island High School traveled to Boeing’s Renton, Washington, facility to learn how planes are constructed and understand what it really takes to have a career building them.

From the outset, the goal was to avoid the typical spectator field trip. Instead, Engineering Tomorrow and Hoffman designed a structured, multi-touch experience that prepared students before they ever set foot in the factory.

First, an Engineering Tomorrow liaison introduced key aerospace concepts and an associated lab challenge to the class via Zoom, then returned in person to guide Mercer students through a hands-on airplane-design lab, helping them translate theory into practice and answer questions about engineering pathways. Students then visited Boeing’s production facility, where they spoke with engineers from multiple disciplines — not just aerospace — and toured the factory floor.

By the time they arrived, students weren’t just impressed by the scale of the operation; they understood what they were seeing, asked informed questions, and left with a sharp sense of the many routes into engineering and manufacturing careers, Dickson says.

“Cameron set up an incredible on-site experience for the students that really made real-world engineering a more tangible experience for them,” Dickson says. “Many people think Boeing is just about aerospace engineering, because Boeing is an aerospace company. But they got to hear from mechanical engineers, electrical engineers, and workers with all sorts of backgrounds who made it clear that there’s no one set pathway into engineering or manufacturing.”

Then came the best part: Students got a VIP tour of the production facility, led by Boeing staff.

A snack and a tour

“It’s awe-inspiring: Dozens of unfinished airplanes are under one site, and you see all of the real-world production engineering that goes into something that oftentimes we take for granted when we step onto an airplane,” Dickson says.

When the big day arrived, students also met with engineering teams to learn about the history of the plant, complete with fun facts geared to high schoolers. (Did you know that a 737 takes off or lands every two seconds?) They learned about different career pathways, from design to production. It was easy to envision themselves working there, Hoffman says.

“Boeing is a company that a lot of folks work at for their entire career and take a lot of pride in the work that they do. We showed them: What does that look like? Do you want to be an engineer for your entire career? Do you want to be a people leader in the facility? Do you want to be a technical expert?” Hoffman says. “And the kids asked great questions.”

Then, the students — after snacks, of course — toured the production floor, where engineers assembled planes and tested parts. For Hoffman, that experience was deeply personal: He wished he’d experienced something similar growing up.

A 10-year Boeing veteran, Hoffman led the group throughout. He started at Boeing in 2015 as a recent college graduate, where he encountered several LGO alums who recommended the program.

“I’d been deeply interested in manufacturing since my early undergrad days. Boeing was an amazing place to work because our products are so complex, and the production systems are so fascinating,” he recalls.

Over time, he wanted to transition into people leadership with an MBA degree. His Boeing colleagues, well-represented among the LGO ranks, urged him toward the MIT program.

“LGO’s network is what makes it so special,” he says.

Upon returning to Boeing after completing his LGO degrees, Hoffman joined Boeing’s LGO/Tauber Leadership Development Program, which allows him to stay regularly engaged with the MIT LGO Program. One such activity where he remains engaged with the program is through the MIT LGO Alumni Board. As part of the board, Hoffman focuses on the social good committee, and the Engineering Tomorrow high school partnership was a perfect fit to meet that committee’s goals.

For Hoffman, these leadership initiatives are what makes LGO distinctive.

“When you graduate from a program like LGO, you’re often so forward-looking. It helps to take time to reflect on what an inspiration you can be to the people who come after you. MIT LGO focuses on both engineering and business. Our students want to study engineering because they want to be problem-solvers. The LGO program, which is at the intersection of engineering and business leadership, is just an incredible inspirational program for young students to see,” Hoffman says.

It was an opportunity he didn’t get as an ambitious young high schooler.

“As a kid, the only engineering class that was available to me was architectural drafting. If this opportunity was offered to me when I was in high school, I would’ve jumped out of my shoes at the chance. You get to see products that are just so complex; you really can't believe it until you see it,” he says.

Setting a positive precedent across industries

Mercer Island engineering design teacher Michael Ketchum had high praise for the field trip, considering it transformative for his students. He estimates that roughly 80 percent of them want to be engineers. He was impressed that the experience was more than just a tour, that it also included classroom support and airplane design kits, reinforcing core engineering concepts. The collaboration allowed them to broaden a previously CAD-focused class into one that also includes 3D printing, electronics, and aerospace applications. 

“For freshmen and sophomores, field trips are key. They stick in their head a bit longer than just school learning. If they get to see people getting excited talking about engineering, and it embeds it a little bit better in their brain,” Ketchum says.

In a post-trip survey, students reported being more likely to consider engineering after the experience.

“They expressed the idea that the conversations with engineers inspired them, and 100 percent of students said that seeing a production facility was one of the coolest parts of the program, which led to them being more inclined to want to be an engineer,” Engineering Tomorrow’s Dickson says.

Next year, the LGO network hopes to expand to partner with additional companies, from health care to biotech.

“The goal is to continue to create exposure. This visit was a really great proof of concept to see what’s valuable to students,” Hoffman says — and, ideally, future LGO alumni.

de MIT News https://ift.tt/yVr5Ddj

Designing a more resilient future for plants, from the cell up

In a narrow strip of land along the Andes mountain range in central Chile, an Indigenous community has long celebrated the bark of a rare tree for its medicinal properties. Modern science only recently caught up to the tradition, finding the so-called soapbark tree contains potent compounds for boosting the human immune system.

The molecules have since been harnessed to make the world’s first malaria vaccine and to boost the effectiveness of vaccines for everything from shingles to Covid-19 and cancer. Unfortunately, unsustainable harvesting has threatened the existence of the tree species, leading the Chilean government to severely restrict lumbering.

The soapbark tree’s story is not unique. Plants are the foundation of industries such as pharmaceuticals, beauty, agriculture, and forestry, yet around 45 percent of plant species are in danger of going extinct. At the same time, human demand for plant products continues to rise. Ashley Beckwith SM ’18, PhD ’22 believes meeting that demand requires rethinking how plants are grown. Her company, Foray Bioscience, aims to make plant production faster, more adaptable, and less damaging to fragile natural supply chains.

The company is working to make it possible to grow any plant or plant product from single cells using biomanufacturing powered by artificial intelligence. Foray has already developed molecules, materials, and fabricated seeds with various partners, including academic researchers, nurseries, conservationists, and companies.

In one new partnership, Foray is working with the nursery West Coast Chestnut to deploy a more disease-resistant version of the chestnut trees that once filled forests across the eastern U.S. but have since been wiped out. The project is just one example of how AI and plant science can be leveraged to protect the plant populations that bring so much value to humans and the planet.

“Plant systems underpin every aspect of our daily lives, from the air we breathe to the food we eat, the clothes we wear, the homes we live in, and more,” Beckwith says. “But these plant systems are fragile and in decline. We need new strategies to ensure lasting access to the plant products and ecosystems we depend on.”

From human cells to plants

Beckwith focused on biology and materials manufacturing as a master’s student in MIT’s Department of Mechanical Engineering. Her research involved building platforms to enable precision treatments for human diseases. After graduating, she worked on a regenerative, self-sufficient farm that mimicked natural ecosystems, and began thinking about applying her work to address the fragility of plant systems.

Beckwith returned to MIT for her PhD to explore the idea of regenerative plant systems, studying in the lab of Research Scientist Luis Fernando Velásquez-García in the Department of Electrical Engineering and Computer Science.

“To address organ shortages for transplants, scientists aspire to grow kidneys that don’t have to be harvested from a human using tissue engineering,” Beckwith says. “What if we could do something similar for our plant systems?”

Beckwith went on to publish papers showing she could grow wood-like plant material in a lab. By adjusting certain chemicals, the researchers could precisely control properties like stiffness and density.

“I was thinking about how we build products, like wood, from the cell up instead of extracting from the top down,” Beckwith recalls. “It led to some foundational demonstrations that underpin the work we do at Foray today, but it also opened up questions: Where are these new approaches most urgently needed? What would it take to apply these tools where they’re needed, fast?”

Beckwith began exploring the idea of starting a company in 2021, participating in accelerator programs run by the E14 Fund and The Engine — both MIT-affiliated initiatives designed to support breakthrough science ventures. She officially founded Foray in February of 2022 after completing her PhD.

“Our early research showed that we could grow wood-like material directly from plant cells,” she says. “We are now able to grow not just wood without the tree, but also produce harvest-free molecules, materials, and even seeds by steering single cells to develop precisely into the products we need without ever having to grow the whole plant.”

Beckwith describes her lab-grown wood innovation as analogous to Uber if there were no internet — a powerful idea without the digital backbone to scale. To create the data foundation and ecosystem to scale plant innovation, Foray is now building the Pando AI platform to enable rapid discovery and deployment of these novel plant solutions.

“Pando functions like a Google Maps for plant growth,” Beckwith says. “It helps scientists navigate a really complex field of variables and arrive at a research destination efficiently — because to steer a cell to produce a particular product, there might be 50 different variables to tweak. It would take a lifetime to explore each of those, and that’s one reason why plant research is so slow today.”

The “operating system for plant science”

Foray’s team includes experts in plant biology, artificial intelligence, machine learning, computational biology, and process engineering.

“This is a very intersectional problem,” Beckwith says. “One of the most exciting things for me is building this highly capable team that is able to deliver solutions that could never be created in a silo.”

After a year of pilot collaborations with select researchers, Foray is preparing for a broader public launch of its Pando platform early this year.

Over the next several years, Beckwith hopes Foray will serve as an innovation engine for researchers and companies working across agriculture, materials, pharmaceuticals, and conservation. Foray already uses Pando internally to create plant solutions that overcome limitations in natural production.       

“Fabricated seeds are one capability that we’re really excited about,” Beckwith says. “Being able to grow seeds from cells lets you create really timely and scalable seed supplies to address gaps in restoration, or shorten the path to market for new, resilient crop varieties. There’s a lot to be gained by making our plant systems more adaptive.”

“We want to shorten plant development timelines, so solutions can be built in months, not decades,” Beckwith says. “We’re excited to be building tools that represent a step change in the way plant production can be done.”

As Foray’s products scale and more researchers use its platform, the company is hoping to help the plant science industry respond to some of our planet’s most pressing challenges.

“Right now, we’re focused on plants in labs,” Beckwith says. “In five years, we aim to be the operating system for all of plant science, making it possible to build anything from a single plant cell.”

de MIT News https://ift.tt/XhqUWrY

New method could increase LLM training efficiency

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.

de MIT News https://ift.tt/85wauBj

Mixing generative AI with physics to create personal items that work in the real world

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.

de MIT News https://ift.tt/tlNbxBG