Rising oil prices are boosting the Kremlin's finances under Russian President Vladimir Putin, but economic growth remains sluggish. Getty Images Soaring oil prices are making Russia richer, even under Western sanctions. High crude prices are boosting Russia's exports, government revenue, and cash inflows. However, labor shortages and weak productivity are keeping growth stuck below 1%. Sanctions-hit Russia is raking in more money from high oil prices, but that doesn't mean its economy is about to take off, according to Goldman Sachs. International benchmark Brent crude futures were trading around $92 a barrel late Tuesday — about 30% higher since the Iran war started. That's a boon to Russia, the world's third-largest oil producer and exporter after the US and Saudi Arabia. Unlike many other oil exporters, Russia doesn't rely on the Strait of Hormuz for oil shipments. That has made the energy giant one of the few beneficiaries of the recent disruption to global oil markets. The problem is that Russia's economy has little room to expand, limiting how much the oil windfall can translate into faster growth, Clemens Grafe, an economist at Goldman Sachs, wrote in a Tuesday note. "Despite the weak growth and funding being available to boost the economy, we do not forecast a demand-driven acceleration," Grafe wrote. Goldman expects Russia's economy to grow just 0.9% this year. That's a slowdown from the 1% and 4.3% growth recorded in 2024 and 2025, respectively. Even so, the country's oil windfall is substantial. Goldman estimates Russia's current-account surplus — a broad measure of trade and income flows with the rest of the world — will nearly double to 3.2% of GDP in 2026 from 1.7% in 2025. Government finances are benefiting, too. Every $10 increase in Russia's oil export price adds roughly $21 billion in budget revenue, according to Grafe. Yet the extra cash isn't translating into faster growth. "There is no meaningful spare capacity in Russia," he wrote. The labor market remains extremely tight, with unemployment near record lows. Productivity growth has weakened, while 2 million workers are no longer available because of military service, casualties, or emigration amid the war in Ukraine, Grafe estimates. As a result, pumping more money into the economy is unlikely to produce significantly more goods and services. Instead, Russian policymakers are likely to view the oil windfall as more of an inflation risk than a growth opportunity. "Higher energy prices will not remove the binding constraints on growth," Grafe wrote. In April, Russian President Vladimir Putin rebuked top officials following the country's economic contraction earlier this year. He demanded proposals for "additional measures aimed at reviving ⁠growth." In January, Putin ordered a "significant increase" this year in tax collection and compliance as the economy came to a standstill. Read the original article on Business Insider

Sophie Steiner has been traveling and dining across China for over a decade. Provided by Sophie Steiner Sophie Steiner has been living in Shanghai for over a decade, and has traveled to nearly every province in China. Her favorite dining cities include Hangzhou, Shangri-La, and Guiyang. In contrast, the subdued flavors she tasted in Nanjing left her disappointed. After a decade of exploring China's culinary landscape — from the shifting-sand deserts of western Xinjiang to the tropical rainforests of southern Yunnan, I've indulged in an ingredient lineup that would fit in the script of "Bizarre Foods with Andrew Zimmern." My journey began in 2015 with a curiosity to explore regional Chinese cuisine. It has evolved into an obsession with the unparalleled gastronomic offerings of this vast country. From street snacks to fine dining, my palate has undergone an epic odyssey. Here are three Chinese epicurean destinations worth the trek, and one that fell flat. Hangzhou, Zhejiang Song's crispy pigeon leg in Hangzhou. Provided by Sophie Steiner About an hour by high-speed rail from Shanghai, where I've been living for over 10 years, Hangzhou is best known for its UNESCO-listed West Lake, tea plantations, and one of China's most refined regional cuisines. Beyond the classics — like braised and lacquered pork, shrimp stir-fried with hyper-local Dragon Well tea leaves, beggar's chicken, and West Lake fish in vinegar sauce — the cuisine encompasses an array of dishes passed down for generations. Wandering the backstreets near West Lake on a recent trip, I came across Wuzi Mianguan, a humble noodle shop with Michelin recognition. Here, hand-pulled noodles are wok-fried with toppings like oil-splash eel, river shrimp, and sautéed pork liver or steeped in a fragrant broth with mustard greens and bamboo shoots. On the high end of the dining spectrum, Song at the Four Seasons Hotel is a sanctuary for all things Zhejiang comfort food, including the signature crispy pigeon leg, laced with a citrusy nip of Sichuan peppercorn heat Shangri-La, Yunnan Dishes in Shangri-La included barley pancakes and dumplings filled with spiced yak meat. Provided by Sophie Steiner In 2020, during my second visit to Yunnan province, I traveled five hours north by train from Kunming to Shangri-La, the high-altitude region along the Tibetan border. Here, the food shifts as dramatically as the scenery. China's most diverse province, home to the country's highest number of minority groups, blends Tibetan, Yunnanese, and Sichuan influences into hearty dishes featuring yak meat and highland barley, shaped by the rugged landscape. Dining in Shangri-La is rooted in community. After a 14-hour hike through village-dotted mountains I was welcomed to my guesthouse owner's family table for bubbling yak hotpot captured Yunnan's warmth and hospitality. My days often began with sweet highland barley pancakes and creamy Tibetan milk tea, followed by spiced yak dumplings for lunch and a communal dinner of cumin-flecked lamb ribs and wild mountain fungi. An unexpected staple is yak-milk cheese, ranging from firm, Parmesan-like wedges to soft curds. It's best paired with a glass of Shangri-La Cabernet. Guiyang, Guizhou Laoguo, a form of dry pot, is served at night markets in Guiyang. Provided by Sophie Steiner Then there's Guiyang, Guizhou's capital, a city rich in minority heritage in southwest China. During my first visit in 2021, I discovered its food strikes the perfect balance of punchy sourness and spice. Sitting on a tiny stool in a smoke-filled room echoing with the clink of Tsingtao bottles, the hiss of sizzling woks, and the cheers of inebriated patrons, I slurped my first sour fish soup — a staple that has become the province's signature dish, and for good reason. The addictively fermented tomato broth is seasoned with floral peppercorn, coriander, lemongrass, and ginger. A buttery river fish simmers within it, surrounded by everything from leafy greens and sliced bamboo to tofu and mushrooms. After sunset, the city transforms into one sprawling night market, with plastic tables and chairs lining Quanlin Pedestrian Street. A nightly feast often involves laoguo, a form of dry pot that pairs perfectly with a frothy beer. While this quintessential street-dining culture is becoming rarer in China's major cities, it's alive and well in Guiyang, drawing me back year after year. Nanjing, Jiangsu Duck blood vermicelli soup in Nanjing was disappointing. Provided by Sophie Steiner In Nanjing, one of China's ancient capitals, the food left me underwhelmed. While the food culture there is sophisticated, with a deep devotion to duck, I found the flavors to be muted. The city's famous salted duck is cured, poached, and served cold, lacking the punch and crunch I crave. Similarly, many rave about the duck blood vermicelli soup, a signature dish of the city, yet the textural symmetry between jelly-like coagulated duck blood cubes and soft vermicelli noodles I find mushy and bland. Nanjing's cuisine emphasizes gentle, savory flavors, but for someone accustomed to the bold spices of other regions, it feels flat and, at times, overly sweet, leaving me longing for the robust tastes I have come to love in other parts of the country. Read the original article on Business Insider

Anthropic released a "Mythos-class" model to the public that includes broad safeguards. Chris Ratcliffe/Bloomberg/Getty Images Anthropic's Claude Fable 5 model has advanced safeguards that can be triggered by mundane requests. The Mythos-class model is blocked or reverts to Opus 4.8 when flagged by safety classifiers. Anthropic said the safeguards were necessary to release the model to the public. If you try asking Anthropic's new Claude Fable 5 model a simple question about cybersecurity or biology, you may find it's not up to the task. That's because the underlying "Mythos-class" model is so powerful that, in order to release it to the general public, it required broad safeguards that can mistakenly flag benign requests, Anthropic said. After some users online said they had triggered the safeguard response with basic prompts about cancer or security, Business Insider put it to the test. I tried asking Fable 5 some simple questions about cancer, like how misinformation about cancer spreads online, and to break down some of the different types. Claude swiftly switched from Fable 5 to Opus 4.8 and notified me of the change before it responded. "Fable 5 has safety measures that flag messages on most cybersecurity or biology topics. They may flag safe, normal content as well. These measures let us bring you Mythos-level capability in other areas sooner, and we're working to refine them," the pop-up said. Anthropic released Fable 5 on Tuesday and said it was as powerful as its Mythos 5 model, only with added safeguards. The release came two months after the company said Mythos was too powerful for a broad release due to cybersecurity concerns. Instead of being released to the public, Mythos was made available only to a small group as part of a cybersecurity project. Anthropic said the safeguards were necessary in order to release the model to the general public. "With the launch of Claude Fable 5, our first Mythos-class model, we believe models now have a greater ability to accomplish real-world scientific tasks and for malicious actors to potentially use our models for highly risky biological research," an Anthropic spokesperson said in a statement to Business Insider. "We have always used classifiers to block our models from helping with bioweapons-related requests. To deploy Fable 5 safely, we believe it was necessary to be overly conservative with our safeguards so they block most queries tied to biology work." Anthropic's Claude reverted to a less capable model when asked basic questions about cancer. Kelsey Vlamis/Anthropic's Claude The company said there are three categories of requests that could get flagged by its safety classifiers: cybersecurity, biology and chemistry, and distillation of Fable 5's capabilities. When the safeguard is triggered, Fable 5 will either be blocked from answering or the model will revert to Opus 4.8 before responding, depending on the user's preference. Anthropic said it was conservative with the safeguards and plans to improve them Anthropic said in its announcement that the safety measures could result in safe, normal content getting flagged, but that their early data showed over 95% of Fable sessions did not fall back to Opus. "To release the model both safely and quickly, we've tuned these safeguards conservatively," Anthropic said, adding it's working on improving the safeguards to reduce false positives. "We intend to make Mythos-class models available without these safeguards to the broader biology and life sciences community so these capabilities can be used to accelerate biomedical research and drug discovery," the Anthropic spokesperson said. The release came about a week after researchers at Anthropic said AI is advancing so fast that frontier labs may need to slow down or temporarily pause so society could keep up. David Kasten, head of policy at Palisade Research, said it's "very clear" from Anthropic's public statements that the company is worried about the risks posed by increasingly powerful models. While he views these safeguards as a good-faith attempt by Anthropic to de-risk, he said historically, "people eventually find a way to get around security restrictions." "It's always a bit of a cat and mouse game between attacker and defender," he said, adding that there is still some risk involved with releasing the more powerful model. He also said that having Anthropic's most powerful model frequently revert to a less capable model could cause a gap in the public's understanding of just how powerful AI models are becoming. "That gap in understanding could be really dangerous for causing policymakers, or for that matter the public, to not fully understand the risks that these models pose in terms of the capabilities they offer," he said. Read the original article on Business Insider

Kalshi, led by CEO Tarek Mansour, is increasing guardrails against insider trading. IMF Kalshi will now check its users' employment status before allowing them to bet. This is the prediction market's latest guardrail to weed out insider trading and bad actors. Prediction markets facing strong scrutiny by lawmakers in the country. Kalshi will now ask you where you work before you can start betting. Kalshi's enforcement and legal counsel, Bobby DeNault, wrote a blog post on the company's website on Tuesday outlining how it is clamping down on insider trading on its platform. DeNault wrote that the New York-based prediction market will start requiring traders to complete employment verification before they can trade, an effort to "screen potential insiders." "This lets us identify presumptive insiders — people who have material, non-public information about a market's outcome — and screen them out before a trade is ever placed," he wrote. Per a screenshot of the employment verification tool embedded in the blog post, Kalshi will ask users for their current employer's name, industry, and job function. Representatives for Kalshi did not respond to a Business Insider query about how long the verification process typically takes, and whether this verification will apply to traders already on the platform. In his post, DeNault outlined two other steps that Kalshi is taking to weed out bad actors. The first is "risk scoring," in which the company assigns a score to each new market added onto its platform, ranging from "niche social or hobbyist markets to critical-scale markets with national or geopolitical reach." This will be done to determine if the market is particularly susceptible to manipulation. And the second is "whistleblower enhancements," which allows users to report suspicious behavior. DeNault called Kalshi users the "natural first line of defense against abusive behavior." These measures work on top of Kalshi's existing guardrails, which include preemptively blocking elected politicians, political candidates, and people in college and professional sports, from trading. Kalshi's new guardrails come as prediction markets like itself, and Polymarket, are facing strong scrutiny from lawmakers in the country, many of whom have proposed bills to ban or stringently clamp down on their activities. For instance, in May, Minnesota became the first state to propose a blanket ban on the markets, with Gov. Tim Walz signing off on it. The US Commodity Futures Trading Commission, which regulates Kalshi and Polymarket, filed a lawsuit against Minnesota and Walz, arguing that Walz cannot criminalize the markets under state law. Read the original article on Business Insider

A member of Ukraine's Typhoon unit holds a Hornet attack drone, known for its AI-targeting capabilities. Nikoletta Stoyanova/Getty Images Mid-range drones are giving Ukraine a way to attack supply routes and command posts in Russia's rear. But drone pilots are also having to spend time testing new entries to the up-and-coming drone class. Some arrive at the front lines riddled with issues, one test pilot told Business Insider. To Spring, some of the new drones blunting Russia's momentum first arrived so glitchy they were practically cursed. "In every sortie, everything that could go wrong went wrong," the drone pilot, with the Ukrainian National Guard's Typhoon unit, told Business Insider of the first system she tested in early 2025. The fixed-wing drone's camera feed would stop working, or its software would stall before takeoff. Controls sometimes became unresponsive after a few minutes, or the batteries failed, she said. And that's after it was already sent to the front for combat use. Spring, identified by her call sign for security purposes, declined to name the specific system. But she said she's tested over 10 types of mid-range strike drones for Typhoon, most of them Ukrainian-made. Their use is growing rapidly on the front line, and more manufacturers are sending in their entries of this drone class for frontline troops to fly. Though considered "finished" drones, their quality can vary greatly, with some encountering severe issues before they clear testing ranges, she said. In combat, such failures can be costly for Ukrainian units, requiring drone pilots like Spring to now juggle time at the front lines with meticulous testing of the fixed-wing systems. Spring poses with a mid-range drone. Her face has been obscured by Business Insider out of concern for her safety. Spring Part of the recipe to Ukraine's success is that these drones, developed at breakneck speed by many manufacturers, are tested by experienced combat pilots, a trial-and-error process that is time-consuming but helps weed out faulty systems. Spring mixes her time on the southern front with several days at practice ranges, conducting up to 11 daily flights that last 30 to 80 minutes each. Drones that start experiencing issues get rigorously tested. "If a manufacturer is not responsible, I do everything possible to prevent their system from reaching combat crews in our unit," Spring said. Troubleshooting while at war Mid-range strike drones are typically classified as fixed-wing systems that can fly between 18 and 180 miles. Analysts say they've provided a critical advantage to Ukraine, allowing it to consistently attack logistics, command posts, and transports in rear areas that Russian commanders had considered safe. Spring mostly specializes in drones that can fly between 40 and 60 km, and has been testing them since early last year. After clearing a drone at a test range, she brings it to the front lines for further trials before it can be deemed combat-ready. Some drones perform well enough on delivery that they only take a few tests for approval, while others are so problematic that they require dozens of flights, she added. A key aspect she assesses is their reliability against Russian jamming after flying long distances. Some mid-range drones are equipped with AI targeting that keeps them flying and continues seeking out their quarry after losing connection to the pilot's control station. Troops prepare a Blyskavka attack drone, a locally-produced fixed-wing system. Anadolu/Anadolu via Getty Images The drones must be able to strike targets reliably at 25 miles away and beyond, a capability that's disrupted Russia's way of fighting and contributed to a net loss of territory for its forces in the last few months. Russia advances by sending waves of foot soldiers against entrenched positions, and any gains they make must be rapidly reinforced with fresh troops, ammo, and artillery. By attacking these assets in the rear, the mid-range drones can hamstring Russian gains. Broadly, the drones that Spring tests cost between $1,000 to $15,000 each, she said. A failed mid-range drone strike not only wastes her unit's money, but can also consume about an hour of flight time due to the distance required to reach the target, she added. There's also a risk that the failed drone falls intact into Russian hands, allowing Ukraine's enemies to reverse-engineer its tech. The cost of a failed strike can be human, too. Ukraine uses many of these drones to hit outposts for Russian first-person-view drone pilots, who in turn are attacking Ukrainian troops. "If the enemy is not stopped, it's a risk to the lives of our people," Spring said. An open-market war industry While testing, Spring said her team stays in close contact with manufacturers and provides feedback on their products. The first mid-range drone she tested, for example, has undergone so many changes since early 2025 that it's now considered a new system. "But that drone is now producing results on the battlefield in the hands of other pilots," Spring said. Mykola Bielieskov, a drone analyst for Ukraine's biggest war crowdfunding organization, ComeBackAlive, told Business Insider that Ukraine's weapons manufacturers compete in a direct-access market with frontline units, where they must maintain their reputations in real time. "The better you respond to data accumulated on the front line, the more guarantees that your UAV will be procured, because units now have separate budgets to procure on their own," Bielieskov said. Mid-range strike drones typically look like small aircraft, with a wingspan of about 6 to 8 feet and roughly the size of a solo kayak. Most are launched by a rail or catapult system and then fly with a propeller on the nose or wings, carrying roughly 20 to 40 pounds of explosives — though there are outliers such as the Fire Point FP-2, which was reported last month to be designed to carry a 440-pound payload for over 100 miles. A member of a Ukrainian drone crew prepares a mid-range drone for launch. Spring One of the best-known mid-range drones in Ukraine is the Hornet, produced by US-based company Perennial Autonomy. Equipped with an AI targeting system, it's supposed to carry an 11-pound warhead for about 93 miles, and was used repeatedly in May by Ukraine's Unmanned Systems Forces to harass Russian logistics highways behind the southern front. Live testing from troops has helped create the rapid innovation cycle that Ukraine is known for. But George Barros, director of Innovation and Open Source Tradecraft at the Institute for the Study of War, said the system also exposes Ukraine's front lines to redundancies from companies competing for profit. Ukraine's open-competition model encourages dispersed production, which has been particularly effective for its FPV drone industry. Roman Chop/Global Images Ukraine via Getty Images "There's a tremendous amount of benefits," Barros told Business Insider. "There's also hidden costs, which are, of course, market inefficiencies, tremendous amount of companies that don't deliver effective or good solutions." Troops like Spring have to spend time filtering them out, though Ukraine also mitigates the issue with a government-run marketplace, Brave1, that rates prototypes based on their reliability. The MRAP 'on steroids' Still, Barros said NATO allies could consider systems that enable troops to gather and provide more direct frontline feedback for manufacturers. The US, for example, introduced the Mine Resistant Ambush Protected vehicle program during the Iraq war as an emergency response to roadside bombs killing troops in Humvees. The MRAP program brought in about a dozen vehicle families for use as troop transports. They're still used today, including by Ukrainian forces. Jose Colon/Anadolu via Getty Images But what Ukraine is doing is the MRAP program "on steroids," Barros said. "That's not something that Americans do. And I think we have to do something closer to that, especially if we were to get involved in a near-peer or a peer adversary war because the adversary will be innovating progressively against us," he said. Gil Barndollar, a non-resident fellow for military analysis at the think-tank Defense Priorities, said Ukraine's "constant, extremely fast feedback loop" benefits from minimal interference from its defense ministry. The US could also learn from that system, said Barndollar, though he cautioned that Ukraine fights on home soil with better access to its industrial base. Washington is building to fight expeditionary wars. "Would we mimic some of their approach in the event of a major war?" Barndollar said. "Perhaps. I think we would move far faster, and perhaps accept more of a 70% solution mindset." "But that line often speaks to Silicon Valley consumer software, not people who have had to send people to war with their products in hand," he added. Translation by Maria Dubarets. Read the original article on Business Insider

arXiv:2606.09853v1 Announce Type: new Abstract: A central objective in multimodal learning is to capture synergy: task-relevant information that arises only from the joint use of multiple modalities, and is not available from any single modality alone. While most approaches operate at the architectural level through larger or more complex fusion models, we propose a complementary axis: shaping the training objective itself. Standard training often emphasizes unimodal or redundant information, falling short on examples that require cross-modal reasoning. We formalize multimodal synergy through information theory and introduce the Synergistic Information Bottleneck (SynIB), a scalable objective that targets synergy directly. To prioritize learning synergy, SynIB motivates the model to predict accurately from all modalities while penalizing confidence when information from any modality is withheld. Alongside the standard task loss, the model runs forward passes with one modality masked at a time and is penalized for remaining confident, which would indicate reliance on unimodal cues rather than cross-modal interactions. We validate SynIB in two regimes. On synthetic XOR tasks where the ground-truth synergy is known by construction, standard training fails to recover it while SynIB does. On five real-world benchmarks, including three MultiBench affective tasks, Hateful Memes with CLIP-ViT and DeBERTa backbones, and a controllable irony extension of CREMA-D we introduce, SynIB improves accuracy on synergy-dependent examples by up to 7.8% and overall accuracy by up to 3.8%.

arXiv:2606.10107v1 Announce Type: new Abstract: Reliable evaluation of instance segmentation models requires metrics that accurately and consistently reflect segmentation quality. However, the metrics most widely used in biological imaging carry fundamental mathematical weaknesses: hard Intersection-over-Union (IoU) thresholds that produce discontinuous, low sensitivity scoring; per-object normalization that distorts scores under object size variation; and greedy or one-to-many matching procedures that yield non-optimal, order-dependent correspondences. Together, these properties produce unintuitive and unreliable model rankings under common failure modes such as split cells, merged cells, and cell boundary imprecision. We propose Maximum Matching Accuracy (MMA), a threshold-free continuous metric that finds a globally optimal one-to-one matching between predicted and ground truth objects and aggregates total overlap using per-pixel normalization. We evaluate MMA against AP@50, PQ, SEG, and AJI across three experiments: synthetic failure cases, progressive corruption tests, and a model ranking comparison. MMA produces scores that are more stable, more sensitive, and more interpretable than existing alternatives, providing a principled foundation for fair instance segmentation benchmarking in biological cell imaging.

arXiv:2606.09856v1 Announce Type: new Abstract: Post-training Large Language Models (LLMs) for reasoning typically focuses on deductive tasks such as mathematics and coding where correctness is verifiable. Yet, many real-world reasoning problems are inductive: agents must infer uncertain beliefs from sparse, ambiguous observations. There are challenges to using standard fine-tuning methods for inductive reasoning, including difficulties in curating large-scale, high-quality labeled datasets and in handling targets that are inherently distributional. In this work, we introduce a novel approach, called Program-based Posterior Training (PPT), to address these limitations: we use an LLM to generate diverse open-world scenarios as probabilistic programs, run probabilistic inference to produce distributional target responses to queries, and then fine-tune on these probabilistic soft labels. Using this approach, we fine-tune LLMs on 10,000 programmatically generated scenarios and evaluate on held-out motifs, human-labeled judgments, and external benchmarks. Overall, PPT substantially improves estimation accuracy on held-out inductive tasks, increases alignment with human judgments, and transfers to external benchmarks for estimation and calibration. Additionally, the gains in raw calibration are not subsumed by post-hoc temperature scaling, showing that the models have more deeply internalized uncertainty compared to output rescaling. Together, these results suggest that probabilistic-program-mediated fine-tuning is a promising approach for post-training LLMs to reliably perform approximate inductive inference.

arXiv:2606.10066v1 Announce Type: new Abstract: Medical vision-language models (VLMs) are evaluated on public benchmarks whose images and question-answer pairs have been freely downloadable for years, yet reported accuracy assumes these examples were absent from pretraining. We audit open VLMs on SLAKE-En, PathVQA, VQA-RAD, and an auxiliary public OmniMedVQA mirror using four detector families: image-side near-neighbour overlap against PMC-OA-beta, canonical-order exchangeability, cohort-relative Min-K%++ tail enrichment, and cross-model top-K overlap. We find measurable image-side source overlap on SLAKE-En: 19.8% of images are flagged under SigLIP-B-16 and 4.2% under SigLIP-SO400M, while out-of-domain controls produce 0/2000 flags. Manual adjudication shows same-modality, same-projection matches to different patients rather than verified pixel-level duplicates, so we interpret this as source or distributional overlap rather than confirmed per-image memorization. On the text side, Qwen2.5-VL on SLAKE-En shows a canonical-order exchangeability signal that survives ordering ablation and external non-medical baselines. On the OmniMedVQA mirror, exchangeability fires for five medical and general VLMs while BLIP-2 remains clean. In contrast, cohort-relative Min-K%++ tail enrichment and cross-model top-K overlap collapse under an external pre-domain baseline: BLIP-2 reproduces the apparent positive signals despite lacking plausible medical-VQA exposure. We conclude that these cohort-relative detectors are unreliable as standalone membership-inference signals on small medical-VLM cohorts.

arXiv:2606.10088v1 Announce Type: new Abstract: Reduced facial expressivity is a common motor manifestation of Parkinson's disease (PD), often described as hypomimia or facial bradykinesia. This paper examines whether temporal motion descriptors extracted from facial-region keypoints can support in-the-wild PD-related video classification on the YouTubePD benchmark. Each video is represented using geometric descriptors from 14 predefined facial regions. Static geometry, normalized geometry, velocity-based descriptors, relative-velocity descriptors, and a GRU sequence baseline are compared under the same binary classification protocol. To assess stability and interpretability, the study includes seed-robustness analysis, region-level ablation, and permutation importance. The best result is obtained with normalized velocity descriptors and a Random Forest classifier, reaching a balanced accuracy of 0.826 and an AUROC of 0.855 on the held-out test split. Across 10 random seeds, this representation remains stable, with balanced accuracy of 0.810 +/- 0.018 and AUROC of 0.855 +/- 0.005. Overall, the results suggest that normalized facial-region motion is a lightweight and interpretable representation for YouTubePD video classification. The study is framed as a benchmark-level analysis and does not claim clinical severity assessment or MDS-UPDRS facial-expression scoring.

arXiv:2606.10021v1 Announce Type: new Abstract: Lumbar spine conditions are a leading cause of disability worldwide, yet reliable quantification of degeneration from MRI remains challenging. In clinical practice, analysis is predominantly performed in two dimensions (2D), as manual three-dimensional (3D) assessment is time-consuming. However, 2D measurements suffer from limited reproducibility, particularly when anatomical structures are not aligned with the imaging plane. Existing automated approaches are often restricted to 2D, rely on discrete grading, or lack robustness and interpretability. We introduce SpineReport, an open-source, fully automated framework for comprehensive 3D morphometric analysis of lumbar spine MRI. Leveraging robust anatomical segmentations, the method extracts quantitative metrics from key structures, including the spinal canal, spinal cord, vertebrae, intervertebral discs, and foramina. These include both morphological and signal-based features, enabling cross-subject and longitudinal assessment. SpineReport further generates subject-specific reports that allow comparison with cohort distributions, improving interpretability and objective characterization of spinal morphology. Clinical relevance was evaluated against radiologist-reported severity grades for central canal, lateral recess, and foraminal stenosis. Metrics showed strong associations with central canal stenosis severity, with T2-weighted CSF signal providing the highest performance (AUC = 0.95). Canal AP diameter and area ratios also demonstrated strong correlations and high discriminative ability (AUC > 0.80). For lateral recess stenosis, associations were moderate, with lateral CSF signal being the most informative (AUC = 0.73). No significant associations were observed for foraminal stenosis despite robust region-of-interest extraction. SpineReport is released as an open-access tool: https://ivadomed.github.io/SpineReport/

arXiv:2606.10019v1 Announce Type: new Abstract: We propose a fast and correspondence-free local point cloud registration method that leverages geometric surface structure and reproducing kernel Hilbert space (RKHS) embeddings. The method represents point clouds as continuous functions with point-wise anisotropic kernels that encode local geometry. This formulation improves alignment along surface normals while relaxing alignment along tangential directions. To solve the resulting registration problem, we propose a second-order on-manifold optimization scheme with approximate Riemannian Hessians, achieving a speedup of up to 10x over the first-order solvers used in prior correspondence-free RKHS-based methods. We demonstrate improved frame-to-frame LiDAR and RGB-D tracking accuracy across diverse indoor and outdoor datasets. On a LiDAR tracking registration task in the driving domain, we achieve a reduction of $>55\%$ in both translational and rotational drift in challenging feature-sparse environments. On object registration benchmarks, we show improved robustness over ICP-based methods and further gains when refining global initialization, particularly under moderate misalignment.

arXiv:2606.09967v1 Announce Type: new Abstract: We present ABot-Earth 0.5, a generative 3D framework designed to synthesize vast, seamless 3D environments from ubiquitous, geospatially referenced satellite imagery. To achieve this, we propose a novel generative model formulated directly with the 3D Gaussian Splatting (3DGS) representation. The model is trained on a diverse corpus of existing real-world urban reconstructions, learning to generate realistic geometry and textures. At inference, it synthesizes novel 3D scenes conditioned solely on satellite imagery at a scalable rate of under 10 minutes per square kilometer, while demonstrating exceptional realism. The framework is designed for accessibility, with integrated hierarchical level-of-detail (LOD) structures that permit real-time, interactive visualization on web-based map engines. This high-fidelity simulation sandbox effectively mitigates the sim-to-real domain gap, enabling critical downstream Embodied AI applications like closed-loop UAV navigation. By providing an ultra-low-cost and high-efficiency solution, ABot-Earth 0.5 significantly lowers the technical and financial barriers to large-scale 3D reconstruction and empowers the future of global digital earth visualization.

arXiv:2606.09882v1 Announce Type: new Abstract: The paradigm of digital twin cities is shifting from coarse visual mapping toward more precise and actionable digitization of urban assets. However, existing datasets predominantly focus on coarse visual perception, lacking the strict multi-modal alignment and attribute and status diagnosis required for automated infrastructure maintenance. To bridge this gap, we introduce WHU-Infra3D, a large-scale, multi-modal benchmark dataset dedicated to roadside infrastructure inventory. Covering 53.8 km across three cities, WHU-Infra3D uniquely integrates panoramic imagery and LiDAR point clouds with rigorous 2D-3D instance association and cross-frame tracking. Comprising over 175k multi-view 2D bounding boxes alongside thousands of 3D infrastructure instances, the dataset provides over 181k detailed attribute and status annotations (e.g., rust, occlusion) to empower operational health assessment. We establish comprehensive baselines across five core tasks: 2D detection, 2D cross-view matching, 3D geo-identification, 3D point cloud segmentation, and attribute recognition. Extensive evaluations expose significant cross-city domain gaps and inherent vulnerabilities of current models on long-tailed defective statuses, establishing WHU-Infra3D as an essential testbed for advancing scalable, AI-driven urban infrastructure inventory and lifecycle management. The WHU-Infra3D dataset is available at https://github.com/WHU-USI3DV/WHU-Infra3D.

arXiv:2606.09871v1 Announce Type: new Abstract: Group Relative Policy Optimization (GRPO) and its variants, originally developed for Large Language Models (LLMs), have recently been applied to Multimodal LLMs and produced strong results. However, their coarse-grained holistic credit assignment from a single scalar advantage underfits vision-language (VL) tasks, where outputs are often long-form responses grounded in semantically rich images. To address this limitation, we exploit a structured signal that single-scalar formulations discard: the natural segmentation of long-form VL outputs. Concretely, we propose Segment-Decomposed GRPO (SD-GRPO), which z-normalizes verifiable per-segment rewards across the rollout group, yielding a vector of per-segment advantages in place of a single scalar. We evaluate SD-GRPO across three settings spanning controlled and real-world long-form VL generation, organized by increasing semantic entanglement across segments. On a controlled multi-panel dense-captioning task constructed from DOCCI, where segments are semantically independent, SD-GRPO consistently outperforms the GRPO baseline, with larger gains at higher segment counts. Extending to a controlled multi-chart long-form VQA task constructed from MultiChartQA, we show both theoretically and empirically that rollout-level rewards suffer from cross-segment credit misattribution that scales with output length. On a real-world scientific figure captioning task on the MMSci dataset, where subfigure captions share context across the figure, blending holistic and per-segment rewards further improves on both, suggesting per-segment normalization alone is insufficient when segments are semantically entangled. Finally, by integrating SD-GRPO into Dr. GRPO, we confirm that it can be applied to any GRPO framework with minimal implementation overhead to enhance long-form VL generation.

arXiv:2606.10126v1 Announce Type: new Abstract: Personalized persuasive text generation can improve relevance and engagement, but demographic conditioning may also introduce unequal framing across groups. We study fairness mitigation in personalized generation as a constrained multi-objective alignment problem: reduce demographic disparities while preserving personalization fidelity. We propose a Pareto-guided teacher alignment framework that combines revision-based candidate generation, pair-aware feasibility gating, Pareto-style candidate selection, and optional preference optimization through supervised fine-tuning and direct preference optimization. We evaluate the framework on climate change and vaccination persuasion tasks using a controlled context-rich demographic grid with matched gender and age pairs and a unified five-audit evaluation suite spanning persuasion bias, formality disparity, emotional framing disparity, lexical association disparity, and personalization fidelity. Across both domains and cross-family transfer settings, no single alignment strategy dominates all objectives simultaneously. Instead, methods occupy different regions of a fairness-personalization Pareto frontier: some achieve stronger disparity reductions, while others better preserve personalization or demographic stability. Our results show that fairness mitigation effects are objective-dependent and transfer inconsistently across domains and model families, motivating bounded-regression, multi-audit model selection over single-metric optimization for fairness-sensitive personalized generation.

arXiv:2606.10113v1 Announce Type: new Abstract: This paper investigates whether LLM translations exhibit identifiable emotional profiles and how post-editing reshapes them toward human-like norms. We compare LLM translations of Margaret Atwood's Oryx and Crake with their post-edited versions and a human translation, using a large-scale corpus of contemporary Italian science-fiction as a baseline. We examine emotion through lexicon-based and multilingual modeling, conducting a fine-grained analysis of emotional variation across systems. We find that MT systems introduce model-specific and statistically significant emotional fingerprints across translations, leading to a limited preservation of an author's voice.

arXiv:2606.10087v1 Announce Type: new Abstract: Pre-training on raw code teaches syntax but provides sparse signal for diverse real-world task formats. While synthetic data has proven transformative for language models, code remains largely unexplored beyond limited quality improvements. We present CodeAlchemy, a synthetic data generation framework that transforms publicly sourced code into semantically-rich training data through 5 strategies: CodeEnhance (quality-aware rewriting), CodeQA (template-based problems), CodeDev (developer tasks), CodeDialogue (multi-turn conversations), and CodeTrace (execution traces). We process 3 corpora across 15 languages to generate 500B+ tokens of synthetic data plus 350B reasoning tokens, orders of magnitude more than prior efforts. CodeTrace instruments and executes 1.3M+ files across 14 languages and 5K libraries, capturing control flow, state tracking, and library knowledge. We introduce DevEval (developer tasks) and TraceEval (execution prediction) benchmarks; frontier models like Claude Sonnet 4.5 achieve only 5.6% exact match on TraceEval, revealing critical gaps in semantic understanding. Our 3B models achieve 83.5% on HumanEval, 63.2% on MBPP, 8.09% win rate on DevEval, and 15.36 ROUGE-2 on TraceEval, outperforming frontier models 10x the size including 27B Gemma-3 and 32B Granite-4.0.

arXiv:2606.10061v1 Announce Type: new Abstract: Large language models (LLMs) increasingly participate in emotionally sensitive social conversations, where responses may shift from balanced support toward excessive validation or escalatory alignment. Existing sycophancy research primarily focuses on factual agreement and instruction-following settings, leaving culturally grounded conversational sycophancy underexplored. We introduce BenSyc, the first benchmark for studying conversational sycophancy in Bengali social contexts. Starting from 11,840 Reddit posts and 170k comments collected from communities across Bangladesh and West Bengal, we construct a human-validated benchmark with binary labels and a fine-grained five-level taxonomy spanning Invalidation, Neutral, Support, Validation, and Escalation. We evaluate more than 15 open and proprietary LLMs on conversational alignment classification and response generation tasks. Results show that distinguishing empathetic support from reinforcement-oriented validation remains challenging even for frontier instruction-tuned models: the best system achieves only 61.8 Macro-F1 on binary detection and 61.7 Macro-F1 on five-class classification. In generation settings, several models frequently produce strongly validating or escalatory responses in emotionally charged situations. Our findings highlight substantial variation across model families and conversational behaviors, underscoring the importance of culturally grounded multilingual benchmarks for evaluating socially aligned conversational AI systems.

arXiv:2606.09900v1 Announce Type: new Abstract: Long-term memory is the missing layer for LLM agents: across sessions they forget, and the common workaround -- replaying the whole history into the prompt -- is expensive, slow, and, as distractors accumulate, less accurate. Most memory systems win on cost or latency but still lose to the full-context baseline on accuracy, and benchmark numbers are reported on inconsistent, non-reproducible harnesses, so one system appears at wildly different scores across sources. We present Engram, an open-source, dual-process memory engine on a bi-temporal data model. A fast write path appends lossless episodes with no LLM on the critical path; an asynchronous path extracts atomic (subject, predicate, object) facts, builds a bi-temporal knowledge graph, and resolves contradictions without an LLM call per fact -- invalidating, never deleting, so every fact keeps provenance and a supersession chain. A hybrid read path fuses dense, lexical, graph, and recency/salience signals, applies a point-in-time ("as-of") filter, and assembles a compact, provenance-tagged context. On the full 500-question LongMemEval_S, graded by the official category-specific judge, Engram's lean configuration -- answering from a ~9.6k-token retrieved slice, never the full history -- scores 83.6% vs. 73.2% for full-context (+10.4 points, McNemar p < 10^-6) at ~8x fewer tokens (9.6k vs. 79k), with 0/500 errored. The gain needs a hybrid read path: facts alone lose recall, while facts plus retrieved chunks recover detail. We also contribute a neutral, in-repo evaluation harness with the official judge baked in and the full-context baseline in every table, publish the raw per-question logs, and document the measurement-integrity pitfalls (truncation, home-grown judges, full-history leaks) that silently distort memory benchmarks. Every number ships with a command to reproduce it.