Committee

arXiv:2602.11424v1 Announce Type: new Abstract: Standard negative log-likelihood (NLL) for Supervised Fine-Tuning (SFT) applies uniform token-level weighting. This rigidity creates a two-fold failure mode: (i) overemphasizing low-probability targets can amplify gradients on noisy supervision and disrupt robust priors, and (ii) uniform weighting provides weak sharpening when the model is already confident. Existing methods fail to resolve the resulting plasticity–stability dilemma, often suppressing necessary learning signals alongside harmful ones. To address this issue, we unify token-level SFT objectives within a generalized deformed-log family and expose a universal gate $times$ error gradient structure, where the gate controls how much the model trusts its current prediction. By employing the Cayley transform, we map the model’s continuously evolving uncertainty onto a continuous focus trajectory, which enables seamless interpolation between scenarios involving uncertain novel concepts and those involving well-established knowledge. We then introduce Dynamic Entropy Fine-Tuning (DEFT), a parameter-free objective that modulates the trust gate using distribution concentration (R’enyi-2 entropy) as a practical proxy for the model’s predictive state. Extensive experiments and analyses demonstrate that DEFT achieves a better balance between exploration and exploitation, leading to improved overall performance.

Americans have a new set of diet guidelines. Robert F. Kennedy Jr. has taken an old-fashioned food pyramid, turned it upside down, and plonked a steak and a stick of butter in prime positions. Kennedy and his Make America Healthy Again mates have long been extolling the virtues of meat and whole-fat dairy, so it wasn’t too surprising to see those foods recommended alongside vegetables and whole grains (despite the well-established fact that too much saturated fat can be extremely bad for you). Some influencers have taken the meat trend to extremes, following a “carnivore diet.” “The best thing you could do is eliminate out everything except fatty meat and lard,” Anthony Chaffee, an MD with almost 400,000 followers, said in an Instagram post. And I almost choked on my broccoli when, while scrolling LinkedIn, I came across an interview with another doctor declaring that “there is zero scientific evidence to say that vegetables are required in the human diet.” That doctor, who described himself as “90% carnivore,” went on to say that all he’d eaten the previous day was a kilo of beef, and that vegetables have “anti-nutrients,” whatever they might be. You don’t have to spend much time on social media to come across claims like this. The “traditionalist” influencer, author, and psychologist Jordan Peterson was promoting a meat-only diet as far back as 2018. A recent review of research into nutrition misinformation on social media found that the most diet information is shared on Instagram and YouTube, and that a lot of it is nonsense. So much so that the authors describe it as a “growing public health concern.” What’s new is that some of this misinformation comes from the people who now lead America’s federal health agencies. In January Kennedy, who leads the Department of Health and Human Services, told a USA Today reporter that he was on a carnivore diet. “I only eat meat or fermented foods,” he said. He went on to say that the diet had helped him lose “40% of [his] visceral fat within a month.” “Government needs to stop spreading misinformation that natural and saturated fats are bad for you,” Food and Drug Administration commissioner Martin Makary argued in a recent podcast interview. The principles of “whole foods and clean meats” are “biblical,” he said. The interviewer said that Makary’s warnings about pesticides made him want to “avoid all salads and completely miss the organic section in the grocery store.” For the record: There’s plenty of evidence that a diet high in saturated fat can increase the risk of heart disease. That’s not government misinformation.  The carnivore doctors’ suggestion to avoid vegetables is wrong too, says Gabby Headrick, associate director of food and nutrition policy at George Washington University’s Institute for Food Safety & Nutrition Security. There’s no evidence to suggest that a meat-only diet is good for you. “All of the nutrition science to date strongly identifies a wide array of vegetables … as being very health-promoting,” she adds. To be fair to the influencers out there, diet is a tricky thing to study. Much of the research into nutrition relies on volunteers to keep detailed and honest food diaries—something that people are generally quite bad at. And the way our bodies respond to foods might be influenced by our genetics, our microbiomes, the way we prepare or consume those foods, and who knows what else. Still, it will come as a surprise to no one that there is plenty of what the above study calls “low-quality content” floating around on social media. So it’s worth arming ourselves with a good dose of skepticism, especially when we come across posts that mention “miracle foods” or extreme, limited diets. The truth is that most food is neither good nor bad when eaten in moderation. Diet trends come and go, and for most people, the best reasonable advice is simply to eat a balanced diet low in sugar, salt, and saturated fat. You know—the basics. No matter what that weird upside-down food pyramid implies. To the carnivore influencers, I say: get your misinformation off my broccoli. This article first appeared in The Checkup, MIT Technology Review’s weekly biotech newsletter. To receive it in your inbox every Thursday, and read articles like this first, sign up here.

Over the past year, Jim O’Neill has become one of the most powerful people in public health. As the US deputy health secretary, he holds two roles at the top of the country’s federal health and science agencies. He oversees a department with a budget of over a trillion dollars. And he signed the decision memorandum on the US’s deeply controversial new vaccine schedule. He’s also a longevity enthusiast. In an exclusive interview with MIT Technology Review earlier this month, O’Neill described his plans to increase human healthspan through longevity-focused research supported by ARPA-H, a federal agency dedicated to biomedical breakthroughs. At the same time, he defended reducing the number of broadly recommended childhood vaccines, a move that has been widely criticized by experts in medicine and public health.  In MIT Technology Review’s profile of O’Neill last year, people working in health policy and consumer advocacy said they found his libertarian views on drug regulation “worrisome” and “antithetical to basic public health.”  He was later named acting director of the Centers for Disease Control and Prevention, putting him in charge of the nation’s public health agency. But fellow longevity enthusiasts said they hope O’Neill will bring attention and funding to their cause: the search for treatments that might slow, prevent, or even reverse human aging. Here are some takeaways from the interview.  Vaccine recommendations could change further Last month, the US cut the number of vaccines recommended for children. The CDC no longer recommends vaccinations against flu, rotavirus, hepatitis A, or meningococcal disease for all children. The move was widely panned by medical groups and public health experts. Many worry it will become more difficult for children to access those vaccines. The majority of states have rejected the recommendations.  In the confirmation hearing for his role as deputy secretary of health and human services, which took place in May last year, O’Neill said he supported the CDC’s vaccine schedule. MIT Technology Review asked him if that was the case and, if so, what made him change his mind. “Researching and examining and reviewing safety data and efficacy data about vaccines is one of CDC’s obligations,” he said. “CDC gives important advice about vaccines and should always be open to new data and new ways of looking at data.” At the beginning of December, O’Neill said, President Donald Trump “asked me to look at what other countries were doing in terms of their vaccine schedules.” He said he spoke to health ministries of other countries and consulted with scientists at the CDC and FDA. “It was suggested to me by lots of the operating divisions that the US focus its recommendations on consensus vaccines of other developed nations—in other words, the most important vaccines that are most often part of the core recommendations of other countries,” he said. “As a result of that, we did an update to the vaccine schedule to focus on a set of vaccines that are most important for all children.”  But some experts in public health have said that countries like Denmark and Japan, whose vaccine schedules the new US one was supposedly modeled on, are not really comparable to the US. When asked about these criticisms, O’Neill replied, “A lot of parents feel that … more than 70 vaccine doses given to young children sounds like a really high number, and some of them ask which ones are the most important. I think we helped answer that question in a way that didn’t remove anyone’s access.” A few weeks after the vaccine recommendations were changed, Kirk Milhoan, who leads the CDC’s Advisory Committee on Immunization Practices, said that vaccinations for measles and polio—which are currently required for entry to public schools—should be optional. (Mehmet Oz, the Center for Medicare and Medicaid Services director, has more recently urged people to “take the [measles] vaccine.”) “CDC still recommends that all children are vaccinated against diphtheria, tetanus, whooping cough, Haemophilus influenzae type b (Hib), Pneumococcal conjugate, polio, measles, mumps, rubella, and human papillomavirus (HPV), for which there is international consensus, as well as varicella (chickenpox),” he said when asked for his thoughts on this comment. He also said that current vaccine guidelines are “still subject to new data coming in, new ways of thinking about things.” “CDC, FDA, and NIH are initiating new studies of the safety of immunizations,” he added. “We will continue to ask the Advisory Committee on Immunization Practices to review evidence and make updated recommendations with rigorous science and transparency.” More support for longevity—but not all science O’Neill said he wants longevity to become a priority for US health agencies. His ultimate goal, he said, is to “make the damage of aging something that’s under medical control.” It’s “the same way of thinking” as the broader Make America Healthy Again approach, he said: “‘Again’ implies restoration of health, which is what longevity research and therapy is all about.”  O’Neill said his interest in longevity was ignited by his friend Peter Thiel, the billionaire tech entrepreneur, around 2008 to 2009. It was right around the time O’Neill was finishing up a previous role in HHS, under the Bush administration. O’Neill said Thiel told him he “should really start looking into longevity and the idea that aging damage could be reversible.” “I just got more and more excited about that idea,” he said. When asked if he’s heard of Vitalism, a philosophical movement for “hardcore” longevity enthusiasts who, broadly, believe that death is wrong, O’Neill replied: “Yes.”  The Vitalist declaration lists five core statements, including “Death is humanity’s core problem,” “Obviating aging is scientifically plausible,” and “I will carry the message against aging and death.” O’Neill said he agrees with all of them. “I suppose I am [a Vitalist],” he said with a smile, although he’s not a paying member of the foundation behind it. As deputy secretary of the Department of Health and Human Services, O’Neill assumes a level of responsibility for huge and influential science and health agencies, including the National Institutes of

The past year has marked a turning point for Chinese AI. Since DeepSeek released its R1 reasoning model in January 2025, Chinese companies have repeatedly delivered AI models that match the performance of leading Western models at a fraction of the cost.  Just last week the Chinese firm Moonshot AI released its latest open-weight model, Kimi K2.5, which came close to top proprietary systems such as Anthropic’s Claude Opus on some early benchmarks.  The difference: K2.5 is roughly one-seventh Opus’s price. On Hugging Face, Alibaba’s Qwen family—after ranking as the most downloaded model series in 2025 and 2026—has overtaken Meta’s Llama models in cumulative downloads. And a recent MIT study found that Chinese open-source models have surpassed US models in total downloads. For developers and builders worldwide, access to near-frontier AI capabilities has never been this broad or this affordable. These models differ in a crucial way from most US models like ChatGPT or Claude, which you pay to access and can’t inspect. The Chinese companies publish their models’ weights—numerical values that get set when a model is trained—so anyone can download, run, study, and modify them.  If open-source AI models keep getting better, they will not just offer the cheapest options for people who want access to frontier AI capabilities; they will change where innovation happens and who sets the standards.  Here’s what may come next. China’s commitment to open source will continue When DeepSeek launched R1, much of the initial shock centered on its origin. Suddenly, a Chinese team had released a reasoning model that could stand alongside the best systems from US labs. But the long tail of DeepSeek’s impact had less to do with nationality than with distribution. R1 was released as an open-weight model under a permissive MIT license, allowing anyone to download, inspect, and deploy it. On top of that, DeepSeek also published a paper detailing its training process and techniques. For developers who access models via an API, DeepSeek also undercut competitors on price, offering access at a fraction the cost of OpenAI’s o1, the leading proprietary reasoning model at the time. Within days of its release, DeepSeek replaced ChatGPT as the most downloaded free app in the US App Store. The moment spilled beyond developer circles into financial markets, triggering a sharp sell-off in US tech stocks that briefly erased roughly $1 trillion in market value. Almost overnight, DeepSeek went from a little-known spin-off team backed by a quantitative hedge fund to the most visible symbol of China’s push for open-source AI. China’s decision to lean into open source isn’t surprising. It has the world’s second-largest concentration of AI talent after the US. plus a vast, well-resourced tech industry. After ChatGPT broke into the mainstream, China’s AI sector went through a reckoning—and emerged determined to catch up. Pursuing an open-source strategy was seen as the fastest way to close the gap by rallying developers, spreading adoption, and setting standards. DeepSeek’s success injected confidence into an industry long used to following global standards rather than setting them. “Thirty years ago, no Chinese person would believe they could be at the center of global innovation,” says Alex Chenglin Wu, CEO and founder of Atoms, an AI agent company and prominent contributor to China’s open-source ecosystem. “DeepSeek shows that with solid technical talent, a supportive environment, and the right organizational culture, it’s possible to do truly world-class work.” DeepSeek’s breakout moment wasn’t China’s first open-source success. Alibaba’s Qwen Lab had been releasing open-weight models for years. By September 2024,  well before DeepSeek’s V3 launch, Alibaba was saying that global downloads had exceeded 600 million. On Hugging Face, Qwen accounted for more than 30% of all model downloads in 2024. Other institutions, including the Beijing Academy of Artificial Intelligence and the AI firm Baichuan, were also releasing open models as early as 2023.  But since the success of DeepSeek, the field has widened rapidly. Companies such as Z.ai (formerly Zhipu), MiniMax, Tencent, and a growing number of smaller labs have released models that are competitive on reasoning, coding, and agent-style tasks. The growing number of capable models has sped up progress. Capabilities that once took months to make it to the open-source world now emerge within weeks, even days. “Chinese AI firms have seen real gains from the open-source playbook,” says Liu Zhiyuan, a professor of computer science at Tsinghua University and chief scientist at the AI startup ModelBest. “By releasing strong research, they build reputation and gain free publicity.” Beyond commercial incentives, Liu says, open source has taken on cultural and strategic weight. “In the Chinese programmer community, open source has become politically correct,” he says, framing it as a response to US.dominance in proprietary AI systems. That shift is also reflected at the institutional level. Universities including Tsinghua have begun encouraging AI development and open-source contributions, while policymakers have moved to formalize those incentives. In August, China’s State Council released a draft policy encouraging universities to reward open-source work, proposing that students’ contributions on platforms such as GitHub or Gitee could eventually be counted toward academic credit. With growing momentum and a reinforcing feedback loop, China’s push for open-source models is likely to continue in the near term, though its long-term sustainability still hinges on financial results, says Tiezhen Wang, who helps lead work on global AI at Hugging Face. In January, the model labs Z.ai and MiniMax went public in Hong Kong. “Right now, the focus is on making the cake bigger,” says Wang. “The next challenge is figuring out how each company secures its share.” The next wave of models will be narrower—and better Chinese open-source models are leading not just in download volume but also in variety. Alibaba’s Qwen has become one of the most diversified open model families in circulation, offering a wide range of variants optimized for different uses. The lineup ranges from lightweight models that can run on a single laptop to large, multi-hundred-billion-parameter systems designed for data-center deployment. Qwen features many task-optimized variants created by the

arXiv:2510.20280v3 Announce Type: replace Abstract: We propose ContextLM, a framework that implicitly learns multi-token prediction by augmenting standard pretraining with an intrinsic next-context prediction objective. ContextLM builds a language model on top of context embeddings that span multiple tokens, enabling better next-token prediction by predicting the next context. Our model is fully compatible with standard autoregressive, token-by-token evaluation paradigms (e.g., perplexity). Extensive experiments with GPT-2 and Pythia backbones (up to 1.5B parameters and 300B training tokens) reveal that ContextLM shifts the Pareto frontier of scaling laws, exhibiting superior efficiency in parameters, training tokens, and FLOPs. Our results show that ContextLM could already achieve the baseline perplexity using 39% fewer parameters and demonstrates robust generalization improvements on extensive downstream tasks under equivalent parameter counts.

EVs are getting cheaper and more common all over the world. But the technology still faces major challenges in some markets, including many countries in Africa. Some regions across the continent still have limited grid and charging infrastructure, and those that do have widespread electricity access sometimes face reliability issues—a problem for EV owners, who require a stable electricity source to charge up and get around. But there are some signs of progress. I just finished up a story about the economic case: A recent study in Nature Energy found that EVs from scooters to minibuses could be cheaper to own than gas-powered vehicles in Africa by 2040. If there’s one thing to know about EVs in Africa, it’s that each of the 54 countries on the continent faces drastically different needs, challenges, and circumstances. There’s also a wide range of reasons to be optimistic about the prospects for EVs in the near future, including developing policies, a growing grid, and an expansion of local manufacturing.   Even the world’s leading EV markets fall short of Ethiopia’s aggressively pro-EV policies. In 2024, the country became the first in the world to ban the import of non-electric private vehicles. The case is largely an economic one: Gasoline is expensive there, and the country commissioned Africa’s largest hydropower dam in September 2025, providing a new source of cheap and abundant clean electricity. The nearly $5 billion project has a five-gigawatt capacity, doubling the grid’s peak power in the country.   Much of Ethiopia’s vehicle market is for used cars, and some drivers are still opting for older gas-powered vehicles. But this nudge could help increase the market for EVs there.   Other African countries are also pushing some drivers toward electrification. Rwanda banned new registrations for commercial gas-powered motorbikes in the capital city of Kigali last year, encouraging EVs as an alternative. These motorbike taxis can make up over half the vehicles on the city’s streets, so the move is a major turning point for transportation there.  Smaller two- and three-wheelers are a bright spot for EVs globally: In 2025, EVs made up about 45% of new sales for such vehicles. (For cars and trucks, the share was about 25%.) And Africa’s local market is starting to really take off. There’s already some local assembly of electric two-wheelers in countries including Morocco, Kenya, and Rwanda, says Nelson Nsitem, lead Africa energy transition analyst at BloombergNEF, an energy consultancy.  Spiro, a Dubai-based electric motorbike company, recently raised $100 million in funding to expand operations in Africa. The company currently assembles its bikes in Uganda, Kenya, Nigeria, and Rwanda, and as of October it has over 60,000 bikes deployed and 1,500 battery swap stations operating. Assembly and manufacturing for larger EVs and batteries is also set to expand. Gotion High-Tech, a Chinese battery company, is currently building Africa’s first battery gigafactory. It’s a $5.6 billion project that could produce 20 gigawatt-hours of batteries annually, starting in 2026. (That’s enough for hundreds of thousands of EVs each year.) Chinese EV companies are looking to growing markets like Southeast Asia and Africa as they attempt to expand beyond an oversaturated domestic scene. BYD, the world’s largest EV company, is aggressively expanding across South Africa and plans to have as many as 70 dealerships in the country by the end of this year. That will mean more options for people in Africa looking to buy electric.  “You have very high-quality, very affordable vehicles coming onto the market that are benefiting from the economies of scale in China. These countries stand to benefit from that,” says Kelly Carlin, a manager in the program on carbon-free transportation at the Rocky Mountain Institute, an energy think tank. “It’s a game changer,” he adds. This article is from The Spark, MIT Technology Review’s weekly climate newsletter. To receive it in your inbox every Wednesday, sign up here.