Financial Times: Sources: audited financial figures show OpenAI spending hit $34B in 2025, including $19B on research and development and nearly $6B on sales and marketing  —  Audited figures reveal mounting costs from model development, infrastructure and rapid expansion

Ashley Stewart / Business Insider: Sources: Microsoft is adding AWS capacity to GitHub after AI-driven growth strained infrastructure and triggered a series of reliability issues  —  Microsoft is turning to its biggest cloud rival, Amazon, to help address capacity issues on its GitHub coding platform following a series of AI-driven outages …

Perplexity CEO shares what he learned from Jensen Huang and Elon Musk about working. Bloomberg/Getty Images Perplexity's CEO said he learned valuable lessons from Nvidia's Jensen Huang and Tesla's Elon Musk. Srinivas said that he doesn't agree with founding a company only to sell it for the money. "You always need to be doing something," he said. "You need to work forever." Perplexity's CEO shared two unusual pieces of advice he learned from the leaders of two of the world's biggest companies. On an episode of the "20VC" podcast released on Monday, Aravind Srinivas said Nvidia CEO Jensen Huang and Tesla CEO Elon Musk taught him two important entrepreneurship lessons that stuck with him. Srinivas cofounded AI search engine Perplexity in 2022 after working as a researcher at Google's DeepMind and OpenAI. In August, Business Insider reported that the company was seeking fresh funding at a $20 billion post-money valuation. The startup's investors include SoftBank, Nvidia, and Jeff Bezos. On Monday's podcast, he said that he learned the importance of always staying on your toes from Huang. "Think about it. $5 trillion, guaranteed to make $500 billion in revenue in the next two years. He has the most advanced chips in the world," he said, about Nvidia. "And he operates with the mentality that he could be 30 days away from going out of business. That is what it takes to be Jensen Huang." Srinivas added that Huang also tells others around him that the chip company is a month away from going out of business. From Musk, he took away the importance of working for more than the money. "If you look at his pay package for SpaceX, it's structured around creating a colony on Mars with a million inhabitants," the Perplexity CEO said. "It's not motivating to be worth 10 trillion in net worth or something." 'Work forever' Srinivas said that he doesn't agree with the entrepreneurship mindset of founding a company, selling it, and then staying home once you have generational wealth. He said that it allows children of founders to have trust funds, but it does not set a good example for them to see their dads sitting at home. "You always need to be doing something," he said. "You need to work forever." His take on entrepreneurship and work contradicts the rapidly growing financial independence, retire early (FIRE) movement. The concept is seen as an ultimate goal by many in the tech community and revolves around retiring in your 30s or 40s after accumulating a net worth sufficient to live off. Shark Tank judge and investor Kevin O'Leary is another vocal opponent of the FIRE philosophy. He retired for a few years after selling his first company and described it as a period when he was "bored out of my mind." "Working is not just about money. People don't understand this very often, until they stop working," he said in a 2019 CNBC interview. "Work defines who you are." Read the original article on Business Insider

Discussion | Link

Australian mining billionaire Gina Rinehart has built major stakes in rare-earth and critical minerals companies alongside her iron-ore business. Philip Gostelow/Bloomberg/Getty Images Australian mining billionaire Gina Rinehart's Hancock Prospecting has invested in SpaceX's record-breaking IPO. Hancock Prospecting says it sees opportunities to work with SpaceX in critical minerals. The IPO raised more than $85 billion and pushed SpaceX's value above $2 trillion. Australian mining billionaire Gina Rinehart has made a major bet on Elon Musk's SpaceX that could lead to more than just a financial return. Rinehart — Australia's richest person with a fortune estimated at $38.4 billion by the Bloomberg Billionaires Index — invested through Hancock Prospecting, her privately held mining and agriculture company. Hancock Prospecting said Monday that it was allocated shares in SpaceX's initial public offering. The company did not disclose the size of its "significant investment," though The Wall Street Journal reported that the stake was worth more than $1 billion. "In the future, we also see the possibility of mutually beneficial arrangements between SpaceX and Hancock Prospecting's significant critical minerals investments, as demand grows for the materials and infrastructure needed to support advanced technology," said Garry Korte, the CEO of Hancock Prospecting, in a statement. He described the SpaceX share allocation as "generous." The comments point to potential opportunities between SpaceX and Rinehart's growing portfolio of critical minerals investments. SpaceX has identified asteroid mining as a potential future opportunity, while NASA is supporting efforts to develop a commercial economy on and around the moon that could eventually include resource extraction. Those ambitions could increase demand for the kinds of critical minerals that Rinehart has spent years investing in. While Rinehart built her fortune through iron ore — the key ingredient in steelmaking — she has spent years investing in rare earths and other critical minerals. Hancock holds major stakes in companies including Australia's Lynas Rare Earths and US-based MP Materials, two of the most prominent rare-earth producers outside China. Rinehart, the executive chairman of Hancock Prospecting, described the SpaceX investment as significant. "We are pleased to have received an allocation in what has been an extremely popular and oversubscribed IPO," she said in the statement. "We see SpaceX as a rare business: led by a truly exceptional person, technically exceptional, and operating in sectors that are crucial, and with long-term potential," she added. The investment gives Rinehart exposure to one of the world's most closely watched technology companies. SpaceX began trading publicly on Friday after raising $75 billion in what was billed as the largest IPO on record. Shares surged nearly 20% in their first day of trading. On Monday, SpaceX announced that the underwriters had exercised a greenshoe option, increasing the amount raised to more than $85 billion. SpaceX closed nearly 20% higher on Monday, lifting its market capitalization to over $2 trillion. Read the original article on Business Insider

The GARA drone is an example of a Ukrainian drone capable of flying for dozens of miles. Equipped with some AI systems, it's a multifunctional platform. Dmytro Smolienko/Ukrinform/NurPhoto via Getty Image Ukraine's new drones are restoring a strike role it once relied on HIMARS to provide. Recent videos show they're beating Russian jamming while carrying heavy explosive payloads. Analysts say the drones have helped Ukraine regain momentum in recent weeks. Three thousand feet above the fields of Zaporizhzhia, a recon drone watches over a deserted village. Its target: a trio of houses surrounded by trees, tucked into a corner of the settlement. All looks quiet until a winged drone flashes into view, skimming over the grass. Built like a small plane, it races toward the largest house, striking the roofline. The explosion is instant. Tiles and debris burst above the trees, and within a second, the upper floor is torn open, smoke and dust pouring from the exposed rafters. "This was a house where Russian FPV drone pilots lived," said Spring, a drone pilot with the Ukrainian National Guard's Typhoon unit. Business Insider reviewed footage of a mid-range drone strike from mid-2025, which she said was her first successful strike. Spring, identified only by her call sign for security purposes, flies a newer type of winged drone that enables Ukraine to consistently strike Russia's rear areas — a capability previously only provided by Western artillery and munitions, such as the M142 High Mobility Artillery Rocket System. Mid-range HIMARS strikes were key to undermining Russia's attack style early in the war, but Russia was able to curb that threat after the first year, analysts told Business Insider. Now, they said, the new drones are bringing that strike effect back in a way Western arms have not been able to do at scale. Drone footage shows the moment Spring sent a mid-range drone flying into her target. Spring "We argue that the Ukrainian mid-range strike is actually heralding a new phase of the war," said George Barros, the director of Innovation and Open Source Tradecraft at the Institute for the Study of War. "What we're looking at here is a really solid foundation for Ukraine to blunt Russian advances." These cheaper, medium-range drones can travel roughly 30 to 300 km, carrying heavier explosive payloads built to devastate command posts, supply trucks, and air defense assets. Some, equipped with artificial intelligence systems, can overcome Russian jamming by autonomously locking onto their target if they lose the pilot's signal. A member of a Ukrainian drone crew prepares a mid-range drone for launch. Spring "In some sectors of the front, they appear to be having a meaningful impact on Russian logistics, which steadily affects front-line forces and makes even the piecemeal Russian infiltration tactics less viable," said Gil Barndollar, a non-resident fellow at Defense Priorities. The use of these drones has risen in the last two months, and Russia has been losing more ground than it has gained, marking a reversal of a yearslong trend in which Ukraine had been slowly bleeding territory. "We're actually quite bullish on the prospects for Ukraine having some substantial upper-hand momentum as we go into the summer," said Barros of ISW. HIMARS supremacy in the war has faded Mid-range drones give Ukraine several new advantages: Kyiv can now conduct intermediate-range strikes at much lower cost, independently decide what to target, and hit those targets more effectively despite jamming. Early in the war, Ukraine's ability to strike deep behind Russian lines was heavily dependent on Western-made weapons. British-French Storm Shadow missiles allowed Kyiv to hit headquarters, rally points, supply depots, and other key rear targets, while its roughly 40 US-made HIMARS launchers could fire rockets up to 150 kilometers and longer-range missiles up to 300 kilometers, delivering similar blows. THe M142 HIMARS was a staple of Ukraine's intermediate-range strikes against Russia. Serhii Mykhalchuk/Global Images Ukraine via Getty Images However, those weapons came with strings attached. Ukraine often relied on donor countries for targeting support, giving foreign governments a say in which attacks could proceed. The US Department of Defense has repeatedly restricted what Kyiv can hit, including a monthslong 2025 freeze on the use of the longest-range HIMARS munitions made available to Ukraine. And by the time the restrictions had been lifted, the weapon's overall effectiveness had been diminished. "The peak of HIMARS was in the summer and autumn of 2022," said Mykola Bielieskov, senior analyst for Ukraine's largest crowdfunding organization, ComeBackAlive. Ukraine only received a limited number of the longer-range missiles, and Russia learned to electronically disrupt the GPS guidance on other HIMARS-launched rockets. As for the Storm Shadows, those are too few to be used at scale. Barros said that by 2023 and 2024, the Ukrainians continued to use HIMARS, but that they were "quite a substantial downgrade" in the face of Russia's jamming. HIMARS was still used from 2023 to 2024, but Russian forces gradually learned to move their logistics in ways that them more survivable. Serhii Mykhalchuk/Global Images Ukraine via Getty Images "In the past, it may have taken one rocket to achieve an effective strike," he said. "Now, it maybe would take four or more rockets to achieve the same level of efficacy." Beating HIMARS at its own game Mid-range drones provide Ukraine with the ability to survive Russian electronic warfare, but the capability wasn't easy to come by. "They have to communicate at this longer-than-100-kilometer range, and then overcome really aggressive Russian jamming," Barros said. "There's a lot of actual technical challenges in being able to maintain communications with the pilot or control stations at that range." Ukraine has worked for years to get that tech to work over longer distances. Now, it's finally showing up on the battlefield, he said. Key to these developments is that artificial intelligence is now reliable enough for many mid-range drones to feature an onboard targeting system that can read what the camera sees and pursue a target in the terminal phase, Barros added. "That is a significant and new technology that is absent in long-range strike tech," he said. A member of Ukraine's Typhoon unit holds a Hornet attack drone, known for its AI-targeting capabilities. Nikoletta Stoyanova/Getty Images For example, the AI can be trained to identify Russian logistics trucks as legitimate targets and attack one even if it loses contact with a pilot or operator. Therein lies another advantage over HIMARS, Bielieskov said. The smaller, more nimble drones can strike moving targets, whereas rocket systems like HIMARS typically attack fixed positions. Bielieskov said that in the last six to eight weeks, there's been video evidence of at least 100 targets, such as fuel lorries, destroyed by Ukrainian mid-range drones in Russia's reserve areas. The full impact on Russia's overall war effort is still a "work in progress," said Bielieskov, adding that it would largely depend on how quickly the Kremlin introduces countermeasures. "The key piece going forward is true autonomy," said Barndollar, the Defense Priorities researcher. "If that can be attained and maintained in an EW-soaked environment, then yes, drones will be replacing most of what artillery does, to include rocket artillery and SRBMs," he said, referring to short-range ballistic missiles. HIMARS-style strikes at scale Another major advantage is the cost. A single Guided Multiple Launch Rocket System, the standard HIMARS munition, costs about $187,000 per rocket, while its extended-range version costs an estimated $479,000 per rocket. The Army Tactical Missile System, the longest-range HIMARS-launched weapon provided to Ukraine, costs about $1 million per missile. Many US partners use HIMARS and its GMLRS and ATACMS munitions, such as Taiwan. Daniel Ceng/Anadolu via Getty Images Mid-range drones, meanwhile, can reach targets at a similar depth while typically costing as little as $5,000 each, though some advanced models are sold for up to $50,000. Spring, the Typhoon drone pilot, said that she's flown about 10 types of mid-range drones, which cost between $1,000 to $15,000. The low-cost allows Ukraine to scale this capability in a way it couldn't with HIMARS or other Western weapons. "We have been seeing more and more footage of these strikes; it means companies like FirePoint are producing at a steady rate," Bielieskov said. FirePoint is a local manufacturer that charges about $50,000 per drone, saying in May that some of its mid-range models would soon carry an explosive payload of 440 pounds. That could put the explosive yield on par, or even above, that of a GMLRS or ATACMS strike, which are about 200 and 500 pounds, respectively. Until then, HIMARS retains an advantage in the pure devastation its payload can unleash, analysts said. Barros said that the US' newest HIMARS munition, the Precision Strike Missile, is much more advanced than the older munitions and would likely still work effectively against Russian jamming. This weapon has not been offered to Ukraine. The US possesses a much more advanced HIMARS munition, the Precision Strike Missile, that is not available to Ukraine and most NATO allies. Official US Army photo by Christopher Bohn "We can't be certain that we can rely on the US," Bielieskov said. "92 kilometers, GPS-guided is not enough, it can be spoofed. So we have the incentive to develop something of a bigger range, more reliable, and with a bigger warhead." Read the original article on Business Insider

Wired: Memo: CTO Andrew Bosworth said Meta did “an atrocious job” sharing Applied AI's vision, promising employees better communication, career growth, and more perks  —  In an internal memo seen by WIRED, Bosworth promised employees more stability, better communication …

arXiv:2606.14943v1 Announce Type: new Abstract: Causal Transformers model sequences through an autoregressive factorization of the joint distribution, which enables efficient left-to-right decoding and conditional likelihood computation. However, they cannot tractably sample from or evaluate arbitrary conditionals -- e.g., a block of text conditioned on past and future tokens. Recent work aims to solve this problem through novel architectures, but they often lead to sub-optimal modeling of such conditionals and degraded generations. We propose Arbitrary Conditionals GPT (AC-GPT) which introduces a simple modification to standard causal Transformers to enable evaluating and sampling from arbitrary conditionals -- including past, future, and mixed contexts -- within a single forward pass. Unlike prior approaches, our method preserves the standard left-to-right ordering and next-token prediction objective essential for both strong performance and efficient training on natural language. Crucially, this compatibility allows existing LLMs to be fine-tuned for arbitrary conditioning. Our empirical results indicate that our method outperforms baselines on modeling arbitrary conditionals, without degrading standard left-to-right performance.

arXiv:2606.14875v1 Announce Type: new Abstract: We study context compression for multi-hop question answering with small language models. We propose Telegraph English, a readable symbolic format that rewrites retrieved passages into structured entity-relation statements, preserving reasoning evidence at lower token cost. In controlled experiments on MuSiQue, TwoWiki, and HotpotQA, Telegraph English outperforms three matched-budget compression baselines (character-level deletion, truncation, and random sub-sampling) on every dataset, with gains of 13 to 20 F1 percentage point. It also outperforms a coherent prose summary produced by the same encoder on the hardest dataset. A pre-registered depth-interaction hypothesis is null: the advantage does not grow with reasoning depth within datasets. We interpret these results as evidence that readable symbolic re-expression preserves entity content more densely than either natural language or coherent summarization at matched token budget.

arXiv:2606.14867v1 Announce Type: new Abstract: Proof autoformalization aims to translate a mathematical informal proof written in natural language into a formal proof in a formal language such as Lean~4. Several works have developed LLM-based models for proof autoformalization. However, existing evaluations have typically focused on translating well-formed informal proofs from curated datasets. We argue that a robust proof autoformalizer must remain faithful even for informal proofs that diverge from these idealized ones, and we present the first study on the robustness of proof autoformalization models. We formulate two categories of perturbations and evaluate robustness under each: a global perturbation paraphrases the informal proof in a different style, under which the formalization should remain consistent; a local perturbation alters a value, symbol, or proof step, possibly in a counterfactual way, and a robust formalization should faithfully reflect the perturbation rather than reverting to the original one or inferring a different one on its own. We build a benchmark with both perturbations on miniF2F and MATH-500, and automatically measure how stable a proof autoformalization's correctness is under global perturbations and how faithfully its output reflects local perturbations. We evaluate seven recent models, all of which are sensitive to global perturbations and mostly fail to remain faithful under local perturbations. Code and data are available via https://github.com/ucr-rai/robust-proof-autoformalization.

arXiv:2606.15017v1 Announce Type: new Abstract: Online web agents often augment a base actor with memory, workflow, or skill modules. These modules can improve performance, but they also consume test-time tokens, a cost rarely reported alongside the actor's inference cost. We study online augmentation, where this overhead is paid on every task, and re-evaluate its benefits under a fixed total inference budget. We compare AWM, ASI, and ReasoningBank with a token-matched vanilla baseline that uses the same budget for additional actor steps. Across three WebArena domains and three models, Gemini 3 Flash, GPT-5.4-mini, and Qwen 3.6-27B, the vanilla baseline matches or surpasses all three augmentation methods in aggregate success rate while often using fewer total tokens. We observe a similar trend on WorkArena-L1 with Qwen 3.6-27B, indicating that the effect extends to enterprise knowledge-work tasks. Our results suggest that skills and workflow memory can be useful in specific domains, but their apparent gains often vanish against a budget-matched actor. We further show that run-to-run variance materially affects outcomes and should be reported as a core evaluation criterion for online web agents.

arXiv:2606.14723v1 Announce Type: new Abstract: We study multiple-choice video question answering on the ImplicitQA benchmark, where the correct answer is never explicitly shown but must be inferred from off-screen events, line-of-sight cues, causal structure, and cross-shot spatial layout. On this benchmark a single frontier video LLM already operates near its accuracy ceiling, and we observe that conventional self-consistency strategies -- majority voting across repeated samples of the same model -- can hurt rather than help, because the model's errors on hard questions are correlated. We propose disagreement-based cross-model routing, a pure inference-time procedure that requires no labels and no training. We triple-sample a native-video model (Gemini 3.1 Pro Preview) at temperature zero, exploit the genuine sample-to-sample variance of its video-processing pipeline to identify the roughly 20% subset of questions where the three samples disagree, and route only that subset to a second model from a different family (Claude Opus 4.8) that consumes uniformly sampled frames with adaptive thinking. On the 1001-question validation set with public ground truth -- our main evaluation -- the method improves AvgAcc by +1.43 over the best single sample of the primary model, with per-category gains concentrated on Motion & Trajectory (+5.49), Inferred Counting (+3.45), and Vertical Spatial Reasoning (+1.82) -- the categories most dependent on cross-shot reference resolution. The same pipeline applied to the held-out 172-question CVPR 2026 ImplicitQA challenge test set achieves 82.03 AvgAcc / 79.71 MacroAvgAcc (+1.81 over the best single sample of the primary model), confirming the validation result on an independent split.

arXiv:2606.14724v1 Announce Type: new Abstract: Video anomaly detection in surveillance settings must balance detection accuracy against real-time throughput, a tension that existing methods address either through stronger feature extractors or more efficient architectures, but rarely both. We present VigilFormer, a unified framework that combines deformable spatio-temporal attention with causal temporal modeling to detect anomalies in untrimmed surveillance video. The proposed Deformable Spatio-Temporal Encoder (DSTE) attends to a sparse set of informative locations across frames, avoiding the quadratic cost of dense attention while retaining the ability to capture irregular motion patterns. A Causal Anomaly Classifier (CAC) applies dilated causal convolutions over snippet-level features and optimizes a contrastive multiple-instance learning objective that separates anomalous and normal representations without frame-level labels. To meet deployment constraints, an Adaptive Confidence Scheduler (ACS) dynamically skips low-information frames at inference time, reducing redundant computation in static scenes. Evaluated on UCF-Crime, ShanghaiTech, and CUHK Avenue, VigilFormer achieves AUC scores of 87.83%, 97.21%, and 89.74% respectively, at 41.5 FPS on a single GPU, outperforming recent weakly-supervised methods in both accuracy and speed.

arXiv:2606.14720v1 Announce Type: new Abstract: This study aims to enhance maritime security by using advanced Artificial Intelligence (AI) and Computer Vision (CV) techniques. For this purpose, it was designed and assessed intelligent object detection systems that can detect the presence of ships on the sea surface under different real-time environments. To achieve this goal, a maritime image dataset with 6,468 images was used, covering different weather conditions like cloudy, foggy, rainy, and sunny environments. Six deep learning architectures were evaluated, including a base Convolutional Neural Network (CNN) model, four transfer learning models (Xception, VGG16, MobileNetV2, and EfficientNetV2L), and a Vision Transformer (ViT) model. The models were compared using multiple performance indicators, including accuracy, Type I and Type II errors, model size, and video processing time. The results show that model performance varies depending on computational constraints and deployment conditions. While lightweight architectures are suitable for resource-limited devices, the ViT achieved the best overall performance, reaching 100% accuracy with the lowest error rates and the fastest video processing time. The findings highlight the potential of AI-driven computer vision systems for maritime surveillance, border protection, and autonomous navigation.

arXiv:2606.15007v1 Announce Type: new Abstract: We introduce Nemotron 3 Ultra, a 550 billion total and 55 billion active parameter Mixture-of-Experts Hybrid Mamba-Attention language model. We pre-trained Nemotron 3 Ultra on 20 trillion text tokens, then extended the context length to 1M tokens, and post-trained using Supervised Fine Tuning (SFT), Reinforcement Learning (RL), and Multi-teacher On-Policy Distillation (MOPD). Nemotron 3 Ultra is our most capable model yet, employing multiple key technologies - LatentMoE, Multi Token Prediction (MTP), NVFP4 pre-training, multi-environment RLVR, MOPD, and reasoning budget control. Nemotron 3 Ultra achieves up to ~6x higher inference throughput as compared to state-of-the-art publicly available LLMs while attaining on-par accuracy. The state-of-the-art accuracy, high inference throughput, and 1M token context length make Nemotron 3 Ultra ideal for long-running autonomous agentic tasks. We open-source the base, post-trained, and quantized checkpoints, along with the training data and recipe on HuggingFace.

arXiv:2606.15026v1 Announce Type: new Abstract: Physiological stress and emotion recognition are important for health monitoring and affective computing. In this work, we present a comprehensive evaluation of deep learning models such as Long Short-Term Memory (LSTM), Temporal Convolutional Networks (TCN), and Transformer on the WESAD dataset for multimodal affect recognition using wrist and chest sensor signals. We perform ablation studies to assess the individual contributions of each modality by training models on wrist-only and chest-only inputs. In addition, we implement a late-fusion ensemble strategy that combines predictions from all three architectures trained on multimodal input. We also employ early fusion at the sensor level by concatenating wrist and chest signals before feeding them into each model. Our results show that Transformer models consistently achieve the highest accuracy in multimodal settings, while TCN models perform best in the wrist-only configuration. The ensemble method yields the highest overall accuracy (98.91 +/- 0.13%) and macro-F1 score (98.56 +/- 0.17%). These findings demonstrate the effectiveness of sensor fusion and ensemble-based fusion in developing robust systems for physiological emotion recognition.

arXiv:2606.14716v1 Announce Type: new Abstract: Edge object detection on embedded hardware requires balancing inference latency and detection quality under changing resource pressure. We present RAMS, a lightweight runtime controller that monitors device pressure, calibrates switching thresholds from idle behavior, and dynamically selects among three resident YOLOv8 tiers (NANO/SMALL/MEDIUM at 320/416/640 px) without model-reload latency. RAMS defines five switching policies, including two detection-conditioned variants that prevent aggressive downgrades after recent vulnerable-road-user (VRU) detections. We further introduce the VRU-Weighted Accuracy Score (SWAS), a scalar metric for offline policy comparison without ground-truth annotations, together with an oracle-bounded variant that separates detector circularity from genuine tier-retention benefit. Across Raspberry Pi 5, x86 laptops, and Jetson Orin ONNX/TensorRT deployments, the same controller equations operate over a 37x latency range. On Jetson Orin TensorRT under heavy load, the safety2 policy achieves 3.41 ms mean latency, 5.6x faster than fixed-MEDIUM inference, while retaining 74% of its proxy accuracy through near-NANO operation with selective SMALL and MEDIUM locks during VRU-positive windows. Detection-conditioned switching improves SWAS by 25.4% under oracle scoring and 47.3% under detector-derived scoring relative to threshold-only policies under heavy load. Live KITTI evaluation reports per-tier VRU recall of 24.2%, 41.2%, and 59.0%, showing that reactive overrides are fundamentally limited by baseline detector recall.

arXiv:2606.14961v1 Announce Type: new Abstract: Chain-of-thought (CoT) reasoning can improve LLM performance, but high answer confidence may be misleading when the accompanying CoT rationale is plausible yet incomplete or poorly supported. We study confidence--rationale alignment: whether a model's confidence in its committed answer is justified by its generated rationale. We introduce a GRPO-based reinforcement learning framework that jointly rewards answer correctness, committed-answer probability, and rubric-based rationale support, where the rubric assesses grounding, coherence, task match, and connection to the selected answer without revealing the gold answer to the judge. Across MedQA, MathQA, and OpenBookQA using three open-weight LLMs, our method reduces the confidence--rationale alignment error by up to 26.51% compared with untuned checkpoints, SFT, and correctness-only GRPO, while maintaining competitive accuracy and often improving calibration. These results show that reliable CoT reasoning requires not only confident answers, but rationales that substantively support them.

arXiv:2606.14725v1 Announce Type: new Abstract: The shift from Convolutional Neural Networks to Transformers has reshaped computer vision, yet these two architectural families are typically viewed as fundamentally distinct. Convolutional Neural Networks are defined by spatially local convolution operations, while Transformers rely on global self-attention. We argue that convolution and self-attention, despite their apparent differences, can be unified within a single k-nearest neighbor aggregation framework. The critical insight is that both operations are special cases of neighbor selection and weighted aggregation. Convolution selects neighbors by spatial proximity while self-attention selects by feature similarity, revealing that they lie on a continuous spectrum rather than representing categorically different computations. We introduce Convolutional Nearest Neighbors (ConvNN), a unified framework that formalizes this connection. ConvNN exactly recovers standard and depthwise convolution by restricting neighbor selection to normalized spatial coordinates, and exactly recovers self-attention and its sparse variants, including KVT-attention, by replacing spatial proximity with scaled dot-product similarity. Beyond these special cases, ConvNN serves as a drop-in replacement for both convolution and attention layers, enabling systematic exploration of the intermediate spectrum between local and global aggregation through configurable similarity functions, neighbor selection strategies, positional encodings, and aggregation kernels.

arXiv:2606.14934v1 Announce Type: new Abstract: This work introduces the Separable Neural Architecture (SNA), a function representational class combining neural approximation with tensor decomposition. The SNA decouples localized coordinate functions (atoms) from global interactions governed by a sparse, low-rank interaction object. This architecture possesses a compact and smooth inductive bias well-suited for solving partial differential equations (PDEs). When viewed as a Galerkin trial space under the variational SNA (VSNA) framework, the formulation satisfies classical variational guarantees under Lax-Milgram: well-posedness, quasi-optimality, convergence, and stability. In high-dimensional spatiotemporal--parametric PDEs, the VSNA mitigates the curse of dimensionality by scaling algebraically rather than exponentially. Exploiting an entirely factorized, tensor-native alternating least squares (ALS) optimization framework reduces this cost to linear in dimension. The VSNA is validated across elliptic, hyperbolic, and parabolic systems, demonstrating close alignment with predicted algebraic and spectral scaling rates. We showcase the SNA as a "solve once, query anywhere" physical world model via two engineering case studies: a 7D parametric manufacturing simulation and an experimental thermal-to-property inversion pipeline for Inconel 718. The VSNA executes a 1,000,000-query Monte Carlo sweep in 102s on a standard laptop CPU, yielding a 150,000x speedup over a full-grid finite element baseline hosted on an NVIDIA A100 GPU. It further enables real-time generative inverse-mode reconstructions under 100ms. These results demonstrate that the SNA serves as a compact mathematical substrate for continuous parameter manifolds to enable real-time inversion, optimization loops, and rapid uncertainty propagation.