Cars queue outside a petrol station in Moscow. Igor IVANKO / AFP via Getty Images Russia is suffering its worst energy crisis in over a decade amid Ukraine's drone strikes. In Crimea, one of the hardest-hit regions, local officials said fuel sales would remain unavailable. At least 78 of Russia's 83 regions are experiencing shortages, with purchase limits in effect in 48. Fuel shortages that have been ramping up since the summer of 2025 are compounding into a nationwide crisis for Russia, forcing the oil-producing nation to ban energy sales and import instead. In the annexed peninsula of Crimea, one of the first Russian-controlled areas hit by the dearth, public gasoline sales were cut weeks ago. "The fuel supply situation remains tense and will persist for some time," Sergey Aksyonov, the Russian-backed head of Crimea, warned again in a statement on Wednesday. "On certain days, fuel will not be available for sale." Aksyonov's announcement comes after Crimea said on June 22 that it was temporarily ceasing fuel sales to everyone but the federal government and essential services. Much of the crisis there stems from a Ukrainian campaign to besiege the peninsula by drone strike. Crimea is connected to mainland Europe and Russia by several important land bridges, all of which have come under intense attack in recent weeks. A new class of longer-range fixed-wing drones, often equipped with artificial intelligence to improve survivability against signal jamming, has enabled Kyiv's forces to hammer these bridges, supply routes, and trucks in the south for months. A key target has been the R-280, a major supply highway running from Rostov-on-Don to Crimea via the western coastline of the Sea of Azov. Traffic on the route dropped over 70%, Ukraine's forces said in June. The strikes continue. On Wednesday, drone pilots from Ukraine's Unmanned Systems Forces said that they had hit more than 360 oil tankers and heavy transport vehicles bound for Crimea in a single week. "Every tanker and every unit of heavy transport on this route is a legitimate military target," the drone warfare branch said in its statement. The fuel problem stretches beyond Crimea. Ukraine has spent the better part of the last year repeatedly bombarding Russian oil and gas infrastructure with long-range, jet-launched drones. Russian government statistics have since said gasoline production fell by 17%, with about a third of local oil refining capacity damaged. Last week, Ukrainian forces said they'd knocked out 42.74% of Russia's oil refining capacity, though this has yet to be corroborated publicly by outside sources. Regardless, the toll is starting to show. At least 78 of Russia's 83 regions are suffering from gasoline or diesel shortages, with 48 of these areas introducing some form of purchase cut or limit, per an analysis by the Institute for the Study of War, a Washington-based think tank. In Moscow, by far Russia's wealthiest city, lines of cars have been seen in recent days outside petrol stations. Social media posts in other regions show queues snaking for miles; one viral video shows a man saying he'd been waiting for 36 hours in Zabaykalsky Krai in the Russian Far East. View this post on Instagram A post shared by νℓα∂ (@vlad_vasheruk) The Kremlin responded on Wednesday by banning diesel and jet fuel exports, saying it had begun importing petrol and sourcing products with "lower environmental standards." Russian leader Vladimir Putin acknowledged the shortages last week, but has so far sought to assure the nation that they will pass, calling them "temporary matters" in a meeting with officials on Wednesday. Read the original article on Business Insider

Financial Times: The AI data center boom has led to surging demand for power transformers, with average lead times for orders, once measured in months, now stretching into years  —  Surging data-centre power demands are intensifying pressure on transformer supply chains.  Transformers are fundamental …

Pixel-Flo, a University of Sheffield spin-out developingmanufacturing technology for MicroLED displays, has raised £5.25 million inseed funding. The round was led by Northern Gritstone, with participa...

Emma Whitford / Inside Higher Ed: A Brown University professor suspected his class used AI to cheat after a take-home midterm averaged 96%, prompting an in-person final, which averaged 48.6%  —  Brown University leaders' response to the alleged cheating incident has been “meek,” the professor said.

Jensen Huang says his engineers love how their role has evolved because of AI. Jung Yeon-je / AFP via Getty Images Jensen Huang said AI is transforming the roles of software engineers at Nvidia. Nvidia CEO has said Nvidia will mass-deploy agents across every division to improve output. Huang rejected the idea that AI simply replaces workers, saying instead that it creates new roles. Jensen Huang says his software engineers love how AI is changing their role. "These agentic systems are new skills, and now we have a lot of software engineers building agents," the Nvidia CEO said in an interview published by the company on Wednesday. "If you ask me, every one of my software engineers prefers to be building agents than to be writing Python code." Huang said that because of AI, Nvidia's engineers are doing less coding, which is like typing. Now, they're working on building agents, benchmarks, and guardrails. "You're taking all the mundane work, and you're trying to get this agent to do it," he said. "That requires imagination, that requires creativity, a lot of technology." Huang, who cofounded the chipmaker in 1993, has been a strong advocate for AI assistants in the workplace. He has repeatedly described a future in which Nvidia will mass-deploy agents across every division to improve productivity. AI agents break down a task into multiple smaller steps, each tackling a specific task to achieve a bigger goal. 'Whole bunch of jobs' In the interview, Huang rejected the idea that AI simply replaces workers, arguing instead that it creates new roles. "The amount of work that we have to do to bring AI into the world is really quite incredible," he said. "So it's creating a whole bunch of jobs. And, my software engineers love this." Unlike Anthropic CEO Dario Amodei and Amazon chief Andy Jassy, Huang has dismissed the increasingly popular concern that AI will mass eliminate white-collar jobs. "This is the part that people don't realize about AI. The first thing that AI is doing right now is creating an enormous number of jobs," Huang said in a May TV interview. "AI creates jobs. AI is the United States's best opportunity to re-industrialize ourselves." Read the original article on Business Insider

Isabelle Bousquette / Wall Street Journal: Companies are mobilizing internal groups of “AI champions” to drive adoption; BCG says 74% of front-line employees now use AI regularly, up from 51% in 2025  —  Companies are increasingly mobilizing internal groups of AI ‘champions’ as a way to sell their more reluctant colleagues on the technology's benefits

Yuji Sugimoto led Bain Capital's acquisition of Toshiba Memory, now known as Kioxia. Shiho Fukada/Bloomberg/Getty Images Kioxia became Japan's most valuable listed company as the AI memory boom sent shares soaring. Bain's dealmaker says South Korea's chip champions benefit from faster, centralized decision-making. He says Kioxia could never have made it this far under Toshiba's ownership. Kioxia may have become Japan's most valuable company amid the AI boom, but South Korea's chip giants still hold a structural edge, according to the Bain Capital executive who led the company's buyout from Toshiba. "The reason South Korean companies are so successful in the semiconductor industry is the powerful top-down leadership and ownership structures of the chaebol conglomerates like Samsung and SK," Yuji Sugimoto, Bain Capital's Japan representative, told Nikkei in an interview published on Wednesday. "In semiconductors, if you can't commit, you fall behind and it's game over. I think it's difficult to manage that kind of business under the governance of large Japanese corporations," he added to the media outlet. Sugimoto led Bain's acquisition of Toshiba Memory, which was spun off from Toshiba in 2017 before being bought by a Bain-led consortium for about 2 trillion yen in 2018. The company was renamed Kioxia the following year. Sugimoto said Kioxia's turnaround would not have been possible had it remained under Toshiba. "It would have been impossible to keep making massive investments while also posting huge losses. Other divisions in the company would have opposed it," he said. Bain's kept investing through the memory downturn even as Kioxia's losses mounted, and the move ultimately paid off as AI-driven demand for memory chips transformed Kioxia's fortunes. "Back in 2018, the term 'AI' was not widely used, and we certainly didn't foresee or fully understand the kind of demand we're seeing today," Sugimoto said. Kioxia's shares have risen more than 4,000% since its December 2024 listing. Bain has since exited its investment, with Managing Partner David Gross telling Bloomberg Television on Thursday that the firm no longer holds a stake in Kioxia. On Thursday, Kioxia's shares were 8.5% higher by midday, helping the Nikkei 225 gain 2%. South Korea's Kospi was 1.7% lower as index heavyweight Samsung Electronics fell 2.3% while SK Hynix traded 2.3% higher. Read the original article on Business Insider

Meta CTO Andrew Bosworth shared new details about a data leak from its AI training program. Bloomberg/Getty Images Meta's technology chief shared details on the data leak from its Model Capability Initiative. Andrew Bosworth said an internal researcher had put the data "in a place it wasn't supposed to go." The program has been paused temporarily since June. Meta CTO Andrew Bosworth shed new light on the data leak that made the company pause its unpopular Model Capability Initiative. In an interview with The Atlantic CEO Nicholas Thompson, released on Wednesday, Bosworth spoke about why Meta paused the AI training program that involved tracking employee keystrokes. The interview was filmed in late June. Bosworth said that data generated by the training program was "quite secure," with only a small number of people having access, but it had been erroneously moved by one of Meta's researchers. "One of the researchers who was working downstream with that data—and there was no breach here— but had put it in a place it wasn't supposed to go," the executive told Thompson. The employee data, in a transformed state, had "landed someplace that it shouldn't have landed internally," he said, adding that Meta did not suspect foul play. The company was "locking the whole thing down" until it could get to the bottom of this incident, Bosworth said. The Model Capability Initiative was introduced in April. It involved installing software on the majority of Meta's US employees to track their keystrokes and mouse movements to train its AI models. The program — and Meta's instruction that employees couldn't opt out of it — drew major backlash from its workforce. Bosworth himself said, during an internal meeting, that employee morale in the company was "probably one of the worst it's ever been" in Meta's two-decade history. However, the program was paused in June after a leak made sensitive employee data accessible to the entire company, according to screenshots seen by Business Insider. "We have carefully designed this program with privacy safeguards, and while we have no indication at this time that any data was improperly accessed by Meta employees, we're pausing it while we investigate," a Meta spokesperson told Business Insider in June. In the interview, Bosworth also shared another reason the program had not gone to plan. It was generating a lot of the same data, he said, when ideally, the company should have gotten more varied data that could be used to train its AI. "Variance is far more important than a high volume of the same thing that gets collapsed into one example, basically," he said to Thompson. "So that was why, a couple of weeks after we initially launched it, we added expanded opt-outs for people who didn't want to do it," he said. "A pause, infinite pause. Whenever you don't want to have it, just press pause." Representatives for Meta declined to provide further comment in response to a query from Business Insider. Read the original article on Business Insider

Bloomberg: Shanghai-based GPU maker Iluvatar CoreX raised ~$902M in a Hong Kong share sale; the company's stock has soared 257% since its January IPO, which raised $473M  —  Shanghai Iluvatar CoreX Semiconductor Co. raised about $902 million from a share sale in Hong Kong after its stock soared since an initial public offering in January.

The Hidden Vulnerability: Why AI’s Connective Plumbing is the Next Security Frontier In March 2026, a … Why AI’s Plumbing Keeps Becoming Its Biggest Attack SurfaceRead more The post Why AI’s Plumbing Keeps Becoming Its Biggest Attack Surface appeared first on .

Gremlin / Getty Images What makes people change their minds, or their behaviour? Social scientists spend a lot of time thinking about this question, and experiments are one of the most powerful ways to answer it. Experiments – testing ideas on real people – take considerable amounts of time and money. Enter large language models (LLMs): artificial intelligence (AI) systems trained to mimic certain kinds of text-based human behaviour based on vast amounts of human-produced text. A new study led by Harvard psychology researcher Ashwini Ashokkumar, published today in Nature, suggests LLMs such as GPT-4 can predict the outcomes of many social science experiments surprisingly well. But the results come with a warning: a system that predicts human responses is not necessarily a system that understands human behaviour, and “synthetic respondents” or “silicon samples” are not a direct substitute for real people. A striking result Ashokkumar and her colleagues assembled 70 real experiments already conducted in the United States involving almost 120,000 participants. They then gave GPT-4 descriptions of hypothetical respondents alongside experimental messages and survey questions, and asked it to estimate how those people would respond under different conditions. Next they compared GPT-4’s predictions with the real results, and found a strong correlation. The model could often distinguish between interventions that were more or less effective. This is a striking result. It suggests LLMs may capture meaningful patterns in the social world – at least in the kinds of text-based US survey experiments examined in the study. But it’s not evidence that AI has discovered a reliable shortcut around human research. Useful forecasts are not the same as understanding US scholars Lisa Messeri and Molly J. Crockett have warned that AI systems can create “illusions of understanding”: outputs that look insightful and useful, while encouraging users to overestimate what has actually been understood. An LLM can generate a plausible explanation or a persuasive forecast. But this may reflect sophisticated pattern-matching rather than genuine insight into the mechanisms behind the observed behaviour. For example, the new study found GPT-4 was often good at ranking the likely effects of different treatments. But it systematically estimated the effects were around twice as big as the real results. That distinction is crucial. A tool may tell researchers that message X will probably work better than message Y, yet still be unreliable about whether the difference will be tiny, modest or transformative. A powerful new tool for pilot studies Used carefully, this information could still be highly valuable. Researchers often run small pilot studies before launching expensive experiments. These pilots help refine interventions and estimate whether a proposed effect is large enough to justify a larger study. LLM-generated forecasts may supplement these pilots. For example, researchers might simulate how different demographic profiles respond to several versions of a vaccination message, workplace intervention or policy framing. However, the new study found that combining LLM predictions with human forecasts was more accurate than either source alone. The most useful future may not be AI replacing human researchers or research subjects, but AI helping researchers decide where to direct scarce human resources. The temptation of ‘silicon sampling’ The idea of “synthetic respondents” or “silicon sampling” is not just of scientific interest. It is increasingly being discussed in polling, market research and public consultation, where supporters see opportunities for faster and cheaper testing. However, critics warn it could undermine trust if simulations are presented as real public opinion. For a politician wondering what the public would think of a new tax policy, or a company curious about a new advertising campaign, LLMs can supply a quick and plausible answer. But it is not quite the same as measuring public opinion. A conventional survey collects responses from people living in a particular society at a particular moment. A synthetic sample draws instead on patterns encoded in the model’s training data, prompt design and guardrails. It may reproduce elements of human judgement, but it lacks lived experience, local knowledge and a real stake in the issue being studied. This gap may be especially important for emerging issues, marginalised communities, fast-moving events, and populations that are poorly represented in online data. Ashokkumar and colleagues found the model performed broadly well across demographic groups, but also identified some differences in accuracy favouring white and Republican samples in the US context. Without careful calibration, synthetic respondents may reproduce dominant patterns in available data while smoothing over disagreement or minority perspectives. Similar applications – and risks – are emerging in AI-assisted expert panels, forecasting and Delphi-style consensus tools. They may make expert forecasting and deliberation processes faster and more accessible, but model-generated agreement can also mask genuine and useful disagreement, rather than resolve it. Silicon sampling may be useful for generating hypotheses and stress-testing assumptions, but it is no substitute for listening to real populations. The risk is not using synthetic agents. The risk is mistaking a model-generated proxy for the population itself. The risk of optimising harmful persuasion The same predictive capability can also be misused. The authors tested whether GPT-4 could identify social media content likely to reduce COVID vaccination intentions. Although the model may refuse to generate anti-vaccination messages directly, it could still help identify which harmful messages from an existing set of options were likely to be most effective. This highlights the need for safeguards that go beyond blocking obviously harmful prompts. Systems may also need protections against using AI to rank, optimise or target harmful persuasion. Research using proprietary LLMs such as GPT-4 is also vulnerable to changes made by model providers. The models can change or be retired without notice, which makes it hard or impossible for other researchers to verify or repeat AI-based findings. The central lesson is not that AI prediction is useless, nor that it is magic. LLMs may become valuable instruments for social science, helping researchers pilot ideas, prioritise interventions and explore scenarios at low cost. At the same time, there’s a risk they may get enough things right to make us think we understand more than we do. Steve J. Bickley is a Research Fellow at Queensland University of Technology and a co-founder of Panalogy Lab Pty Ltd, in which he holds an equity interest. Panalogy Lab develops SurveyLM, an agent-based simulation platform for social and behavioural research. He has previously received research funding and institutional support for work involving synthetic agents and AI-assisted experimental research. His relationship with Panalogy Lab is managed under QUT’s approved Conflict of Interest Management Plan.

Gu Yichen was laid off shortly after starting at Google and discovered how unstable life on an H-1B visa can be. Provided by Gu Yichen Gu Yichen left Amazon for what he thought was his dream job at Google. His entire team was laid off before he did a single day of real work, leaving him in a race to save his H-1B visa. He returned to Amazon and says the setback taught him that career success depends as much on timing as effort. This as-told-to essay is based on a conversation with Gu Yichen, a 31-year-old Chinese national who lives in California. After being laid off by Google, he returned to Amazon, where he works on an H-1B visa. His words have been edited for length and clarity. I spent my sophomore year of high school as an exchange student in Yacolt, a small town in Washington State. I didn't realize that participating in the program would lead me to skip China's entrance exam and set me on a path toward studying abroad for my bachelor's. This is part of our ongoing coverage of how the Trump administration's H-1B visa changes affect workers. If you've been impacted and want to share your story, please fill out this short form for the reporter. I also hadn't realized that it would eventually lead me to build a career in the US. In his teens, Gu traveled to the US as an exchange student and lived in a small town in Washington State. Provided by Gu Yichen Getting the H-1B lottery I majored in electrical and computer engineering at the University of Illinois Urbana-Champaign. My focus was on circuit boards and chips at the time, but I was more interested in the computer side, because results were more tangible and visible. During my junior year, I interned at Amazon. After graduating, in 2017, I started my first full-time job on that same team, while on an OPT visa. In college, he got an internship at Amazon. Provided by Gu Yichen Having a STEM degree gave me three chances to enter the H-1B lottery. I got lucky on my third try. I had to return to China to get my visa, and because of COVID, I ended up working remotely for Amazon for a year. Laid off abruptly I successfully applied for a job at Google in late 2022. Companies were hiring aggressively, interviews were relatively easy, and compensation packages were huge. My manager at Amazon tried to convince me to stay, saying things were unstable and the future was uncertain. I felt that if I didn't take the chance while I was young, I'd be less likely to do it later. Human resources reassured us at orientation that there wouldn't be layoffs. My team had planned to work on an experimental project. Due to cost-cutting, it was shelved, and the entire team was let go. I had started work around Christmas, and the layoff notice came in January 2023. I didn't do a single day of real work. My former manager was right. He returned to China after losing his job at Google. Provided by Gu Yichen Slow life in Yunnan I reached out to friends at Google to see if their teams had vacancies, but I couldn't find anything. There was also a time crunch because I was on an H-1B visa: Within a 60-day grace period, I had to find a job, get sponsorship transferred, and start working. It was tight. I felt companies were laying people off everywhere. If I applied, I'd likely end up in a stopgap role I didn't see myself staying in long term. I decided to take a break instead. I went back to my hometown, Nanjing, for a while, then traveled to Yunnan province in southwestern China to stay with my aunt. Return to the US I'm not the type who can rest for long. I was hoping that, as a Google alum, I'd be rehired. If a position became available within six months of leaving, I wouldn't have to go through the interview process again. I was also still in contact with colleagues from my old team at Amazon. They told me a position had opened up. I could continue using my previous H-1B application rather than starting from scratch. If I switched to a new team or company, I'd have to restart the application, which has become harder in recent years. Because I returned to the same team at Amazon in Sunnyvale, California last year, recent H-1B changes under the Trump administration haven't affected me. Gu enjoys the work environment in the US. Provided by Gu Yichen I've applied for a green card, and once my I-140, a key approval in the process, is approved, I can renew my H-1B indefinitely. It will give me more freedom in both my personal life and my career. In the future, I might start my own business or take cooking lessons. The work environment in the US feels like a better fit for me. As long as you get the job done, nobody cares whether you work during the day or at night. There's no drama. My experience at Google made me realize that while I prefer working on experimental projects, companies tend to prioritize essential teams and mature products over exploratory projects. It also taught me that landing your ideal job is often as much about timing as effort. I became more flexible once I realized how much was outside my control. Read the original article on Business Insider

Meta's data center facility in Alberta, Canada, is expected to bring 300 operational jobs, the company said. Meta Meta is building a 1 gigawatt "AI-optimized" data center in Alberta. It would be Meta's first data center in Canada and its 33rd facility worldwide. Meta said it's investing $9 billion in the site and another $40 million in local infrastructure. Meta is bringing a data center to Canada. The company announced on Wednesday that it broke ground for a new AI computing center in Sturgeon County, Alberta, just outside Edmonton. The facility represents an investment of more than CAD $13 billion, or roughly $9 billion, and is planned as a 1-gigawatt data center, the company said in a press release. It will be Meta's first data center in the country. "This data center will be optimized for our AI workloads, helping bring to life the technologies that billions around the world use to connect, find communities, grow businesses, and experience the power of our wearables," Meta said. The facility will be Meta's 33rd data center globally. Meta said it expects the project to support more than 3,000 construction workers at peak and more than 300 "operational jobs" once it's completed. The company said it will also spend about CAD $60 million, or about $42 million, on local infrastructure improvements. A Meta spokesperson told Business Insider that the data center will cover 2.9 million square feet and occupy 1,750 acres. "The design of this data center gives it the ability to support the next generation of AI systems — everything from the hardware equipment and network design to the cooling systems," the spokesperson said. The company said the facility will use a "closed-loop, liquid-cooled system with dry cooling," meaning it is designed to avoid the need for a continuous water supply for cooling. The spokesperson said Meta anticipates the facility will come online in the next 2 to 3 years. In a statement, Alanna Hnatiw, the mayor of Sturgeon County, welcomed the new facility as a positive development for the region, bringing jobs and "long-term tax revenue." "We're excited to work with our new neighbors as we continue to make that vision a reality," the mayor said. In the US, Meta said it has 28 data centers, including the massive Richland Parish site in Louisiana. The Louisiana data center campus is expected to be Meta's largest facility to date, with a footprint of 4 million square feet, providing more than 2 gigawatts of compute capacity. Read the original article on Business Insider

Laptops and headphones in a classroom in Texas. Bloomberg/Getty Images AI isn't just shaking up workplaces and college classrooms; it's gaining ground in grade schools. Half of parents in a recent survey said they were concerned their child "relies on AI too much." At the same time, one-third of parents say schools aren't teaching kids enough AI skills. Artificial intelligence really is everywhere these days. Beyond shaking up workplaces and college classrooms, the tech is gaining ground among grade school students — and it's making some parents nervous. Among 1,150 parents of school-aged children surveyed in Deloitte's annual back-to-school survey, half said they were concerned their child "relies on AI too much." That usage is outpacing the number of parents who said their child's school provides approved generative AI tools (22%) or has established guidelines for using the tech (33%). Nearly 30% of respondents also said their children were already using generative AI tools in their schoolwork. At the same time, more than a third of parents said they are concerned schools aren't preparing kids enough with AI skills, with one in eight parents planning to pay for AI tutoring or camps. The numbers add a new wrinkle to the growing debate over tech tools in schools. Business Insider's Katie Notopolous wrote in May about how her third grader and his friends "had been using Google's Gemini on their school-provided Chromebooks to make funny pictures of poop and dinosaurs." Several school districts are also balancing a backlash against the widespread use of video resources like YouTube against a yearslong decline in math and reading scores. One physics teacher in Canada told Business Insider last year that students' use of AI tools had prompted him to include more analog classwork. "I've tried to sort of shift back toward some handwritten assignments, instead of having them do it on the computer," Ward said. "That way, I can tell this is how they're writing. I know it's theirs." Read the original article on Business Insider

arXiv:2607.06845v1 Announce Type: new Abstract: African American English (AAE), a rule-governed dialect spoken by over 30 million people, is routinely misinterpreted and "corrected" by large language models (LLMs). Across six instruction-tuned LLMs (14B to 70B), we show that state-of-the-art models systematically prefer Standard American English (SAE) continuations even when the preceding context is in AAE, effectively rewriting AAE into SAE. We present an end-to-end framework to audit and mitigate this bias. For auditing, we introduce conditional Dialect Group Invariance (cDGI), which isolates true model bias from translator-induced artifacts, and a feature-level localization analysis that identifies which AAE markers most strongly trigger bias; we find that syntactic constructions, especially negative concord (e.g., "ain't nobody"), are universal triggers across all models. For mitigation, we introduce, to our knowledge, the first application of activation steering to dialect bias: a training-free, test-time method that extracts dialect directions via causal tracing and injects them into bias-relevant layers. Activation steering reduces bias 5 to 20 times more than prompting while preserving SAE fluency. To enable this work, we release REAL-AAE , the largest real-AAE parallel corpus to date: 17,479 AAE/SAE/ AAE_back triplets from natural tweets (2 to 6 times larger than prior real-AAE resources), validated automatically (BERTScore F1 = 0.95) and by three native AAE speakers (83.0% semantic agreement).

arXiv:2607.06587v1 Announce Type: new Abstract: We present CoFINN (Conservation Flux Informed Neural Networks), a physics-informed deep learning framework for predicting compressible flow fields governed by conservation laws. Unlike conventional data-driven convolutional neural networks (CNNs), which optimize only pixel-wise similarity metrics, CoFINN embeds finite-volume conservation physics directly into the training process. Unlike classical physics-informed methods which enforce differential-equation residuals at collocation points through automatic differentiation, CoFINN adopts a finite-volume perspective consistent with modern CFD methodology. CoFINN interprets CNN output fields as structured computational grids, where each pixel represents a finite-volume cell, and enforces conservation consistency through sophisticated numerical flux calculations. The framework is evaluated on transonic flow prediction around airfoils at (M=0.7, Re=6 * 10^6), including challenging conditions involving shock waves and high angles of attack. Results show that CoFINN improves aerodynamic force prediction accuracy, reducing drag prediction error by up to 34% at extreme angles of attack and by approximately 15% on average across the test set. Improvements are particularly significant in limited-data regimes, demonstrating that the conservation-based loss acts as an effective physical regularizer. The proposed approach maintains the computational efficiency advantages of CNN surrogates while significantly improving physical consistency and conservation behavior. The framework is architecture-agnostic and extensible to broader classes of conservation-law-governed physical systems.

arXiv:2607.06616v1 Announce Type: new Abstract: Label skew in federated learning (FL) causes client drift and degrades global accuracy. Synthetic data augmentation can reduce this imbalance; however, full class balancing requires substantial computation cost. We propose FedEAS, a policy that assigns each client an entropy-adaptive per-class generation budget computed from its local label distribution. The budget jointly decides \emph{how much} each client generates and \emph{WHERE} the samples go. Accordingly, the total generation budget follows from the per-client budgets rather than being fixed in advance. FedEAS recovers most of the accuracy gain of full class balancing while reducing the generation budget by 94.1\%. At the same total generation budget, it outperforms Uniform allocation by up to 18.82\% across CIFAR-10 and CIFAR-100.

arXiv:2607.06614v1 Announce Type: new Abstract: Accurate station-level demand forecasting is essential for the efficient operation of bike-sharing systems, yet it remains challenging due to complex spatio-temporal dependencies and the large scale of urban networks. This paper presents STAGformer, a Spatio-Temporal Agent Graph Transformer that achieves efficient global modeling with linear computational complexity. The model introduces a two-step agent attention mechanism, where a small set of learnable spatial and temporal agent tokens first aggregate global information and then broadcast it back to individual stations and time steps, effectively capturing long-range interactions while reducing the quadratic cost of standard self-attention to O(NT). STAGformer integrates four core modules: a spatio-temporal encoder that fuses dynamic node features with external contextual factors (weather, time, points of interest), a graph propagation module for spatial neighbor aggregation, a temporal convolution module for local pattern extraction, and the agent attention module for global dependency modeling. Extensive experiments on two real-world datasets -- NYC Citi-Bike and Chicago Divvy-Bike -- demonstrate that STAGformer consistently outperforms state-of-the-art baselines across multiple prediction horizons, achieving significant improvements in both RMSE and MAE. Ablation studies validate the contribution of each component, with the agent attention mechanism proving critical for modeling global spatio-temporal dependencies.

arXiv:2607.06607v1 Announce Type: new Abstract: Accurate long-term forecasting in complex systems is frequently compromised by dataset-level distribution shifts, where diverse underlying behavioral modes and evolving system states drive the dynamic multivariate time-series. While existing methods predominantly focus on local temporal shifts, they fail to explicitly model the global structural challenge where datasets are composites of distinct operational regimes. In this paper, we propose NEST, a specialized framework designed to model and recompose these evolving structures through a two-phase dense MoE architecture. NEST first facilitates structural specialization by partitioning the dataset into distinct operational regimes through unsupervised clustering in a principled moment-entropy space. We introduce a regime-oriented router mechanism that generates initial expert weights based on temporal content, subsequently refined through geometric modulation to regime centroids. Crucially, rather than acting as monolithic predictors, individual experts function as specialized kernels that capture regime-specific dynamics by evolving unique variate-attention patterns. Extensive evaluations on diverse benchmarks, including heterogeneous network traffic and physical phenomena, demonstrate that NEST consistently achieves state-of-the-art performance. Our code and datasets are available at https://github.com/Aaralshin/NEST

arXiv:2607.06603v1 Announce Type: new Abstract: The notion of algorithmic fairness has been actively explored from various aspects of fairness, such as counterfactual fairness (CF) and group fairness (GF). However, the exact relationship between CF and GF remains to be unclear, especially in image classification tasks; the reason is because we often cannot collect counterfactual samples regarding a sensitive attribute, essential for evaluating CF, from the existing images (\eg, a photo of the same person but with different secondary sex characteristics). In this paper, we construct new image datasets for evaluating CF by using a high-quality image editing method and carefully labeling with human annotators. Our datasets, \oursceleb and \ourslfw, build upon the popular image GF benchmarks; hence, we can evaluate CF and GF simultaneously. We empirically observe that CF does not imply GF in image classification, whereas previous studies on tabular datasets observed the opposite. We theoretically show that it could be due to the existence of a latent attribute $G$ that is correlated with, but not caused by, the sensitive attribute (\eg, secondary sex characteristics are highly correlated with hair length). From this observation, we propose a simple baseline, Counterfactual Knowledge Distillation (CKD), to mitigate such correlation with the sensitive attributes. Extensive experimental results on \oursceleb and \ourslfw demonstrate that CF-achieving models satisfy GF if we successfully reduce the reliance on $G$ (\eg, using CKD).