OpenAI has launched a Partner Network backed by a $150 million investment. This initiative aims to support global partners in accelerating the adoption of enterprise AI. The funding will focus on helping businesses with AI deployment and transformation efforts. The goal is to further drive worldwide utilization of advanced artificial intelligence solutions.
Graham Platner's victory in the Maine Democratic primary suggests voters are more attuned to economic grievances than simply rejecting incumbents. Research on over 36,000 voters across multiple countries reveals a notable opposition to corporate power and billionaire influence. This pattern indicates that general dissatisfaction stems from perceived systemic rigging by large corporations. The findings suggest that anti-establishment sentiment focused on economics is a significant current voting trend. Such evidence points toward voters seeking change in the economic structure rather than just targeting specific political figures.
Andrew Hastie compared the current state of artificial intelligence development to the Cold War nuclear arms race. He warned that without significant investment, Australia risks losing its strategic independence. Hastie cautioned that the nation could become a "supplicant state" overly reliant on the US amidst potential conflict with China. As such, he urged substantial scaling up of Australian AI investments. His proposal was for Australia to proactively position itself as a major technology hub in the southern hemisphere.
Europe is developing digital sovereignty plans amid concerns over dependence on US big tech companies. Personal examples, like those of Slovenian judge Beti Hohler, illustrate how sanctions from US entities can instantly cut off access to essential services such as payment systems and travel platforms. These experiences create a state of "constant uncertainty" for Europeans dealing with geopolitical risks impacting digital access. While aiming for self-sufficiency, the author suggests that current European efforts may still adhere too closely to established Silicon Valley norms.
Anthropic has withdrawn its image generation models, Mythos and Fable. This action was prompted by a directive from the United States government. The withdrawal signifies changes in Anthropic's approach to AI-generated visual content. Previously, these tools allowed users to create images within their applications. Users can now find an alternative service offering to transform ChatGPT images into editable Canva designs.
Entering the Artificial General Intelligence (AGI) era represents an irreversible shift in technology. This development significantly challenges existing frameworks of AI governance. Current governance structures were unprepared for the magnitude of this transition. The rapid advancement suggests a critical need for updated regulatory approaches. Addressing this gap is paramount due to the transformative and irreversible nature of AGI.
The authors argue against the narrative of mass job displacement due to AI, even within software engineering. Evidence from state unemployment filings shows no correlation between AI and layoffs in recent reports. They assert that coding speed is not the primary bottleneck in software engineering; rather, tasks like meetings or debugging are more complex areas. The true bottlenecks identified are deciding what to build, verifying deliverables, and possessing deep human understanding of both the business context and the codebase. Ultimately, providing value depends on this unique "deep human understanding," which AI assistance cannot fully replace.
Small businesses are crucial to sustaining the tandem bike industry. A California-based small company successfully popularized tandems throughout the United States, managing to persist despite larger industry competitors. The piece explores the role of these smaller enterprises in keeping the niche market alive. It highlights their importance against larger corporate players in the cycling sector.
Flash-KMeans is an open-source implementation of standard Lloyd's k-means running on Triton GPU kernels. It maintains mathematical exactness while improving performance by eliminating distance-matrix materialization via FlashAssign and reducing atomic contention with Sort-Inverse Update. Benchmarking shows it achieves significant speedups, reporting 17.9$\times$ end-to-end speedup compared to cuML. Furthermore, it demonstrates an impressive speed improvement exceeding 200$\times$ when benchmarked against FAISS on an NVIDIA H200 GPU.
Z.ai released GLM-5.2 on June 13, 2026, making it available across all GLM Coding Plan tiers. Key features include a usable 1-million-token context window and two adjustable thinking effort levels (High and Max). Access is provided through an Anthropic-compatible endpoint for tools like Claude Code, Cline, and OpenClaw. Notably, the initial launch did not include published benchmarks. Z.ai also committed to releasing MIT open weights next week.
Claude Code functions as a layered agentic coding tool rather than a simple chat prompt. The guide details twenty-five specific features, ranging from CLAUDE.md and skills to subagents, hooks, MCP, and Auto Mode. It provides readers with actionable content, including a comparison table and practical working code examples. Furthermore, the guide offers real-world use cases to illustrate how these features can be implemented. An interactive demonstration is also included for users to experience the tool firsthand.
A hands-on tutorial explores the massive FineWeb dataset without requiring a full download. The workflow demonstrates streaming a manageable sample of the corpus while inspecting its schema and metadata. Key analysis focuses on fields like URL, language, language score, and token count. Furthermore, the guide reproduces simplified versions of FineWeb's quality-filtering pipeline. This comprehensive process addresses large-scale web corpus analytics tasks, including filtering, deduplication, and tokenization.
Vision Large Language Models (LLMs) extend traditional document parsing capabilities by interpreting visual elements within documents, in addition to extracting text. Unlike other parsers that only read written words, vision models can process images such as charts and diagrams found in PDFs. This capability is particularly valuable for Retrieval-Augmented Generation (RAG) systems. By understanding graphical data alongside text, these models provide a richer context for information retrieval. Effectively, vision LLMs transform them into comprehensive PDF parsers capable of handling multimodal inputs.
Examines the underlying microarchitectural expenses associated with implementing GPU time-slicing within a Kubernetes environment. Specifically investigates the practical overhead of co-locating multiple Agentic AI workloads utilizing shared GPUs. The discussion focuses on deep, systems-level performance analysis rather than high-level implementation guides. Understanding these costs is crucial for optimizing resource allocation for concurrent LLM agents.
A new reinforcement learning practitioner created an agent to play the game Little Cricket Master. This project served as a major learning experience for the author's first venture in Computer Vision and RL. The developer has made progress with the agent, noting it can currently score half centuries. Further refinement of the reward function is planned for improvement. A repository containing the code was provided for viewing.
Details the manual tracing of multi-agent self-correction failure modes and context window inflation issues. The analysis method explicitly avoids the use of wrapper frameworks, indicating a deep dive into underlying mechanisms. This hands-on approach aims to provide granular understanding of how these failures occur. Specifically, it diagnoses problematic behaviors within complex agent interactions. The work represents a detailed, low-level examination of system limitations.
An MSc Artificial Intelligence student is seeking guidance to build a strong portfolio for job applications after graduating with a thesis on MuJoCo Gym and World Models. The post outlines desirable candidate profiles for roles in the field. Ideal candidates possess an MSc in relevant areas, hands-on experience deploying RL agents beyond simulation, and deep knowledge of modern RL/MARL techniques like PPO and SAC. Required skills include strong Python/C++/Rust coding ability and production experience with deployment pipelines. The opportunity encourages applications from underrepresented groups despite not meeting every listed criterion.
A user on a forum thread is questioning others regarding unpredictable outcomes when conducting multiple separate AI coding sessions. The post appears to be a general query for community discussion. No specific details about the messy results or the nature of the coding problem are provided in the visible content. The interaction structure suggests seeking advice or warning from fellow developers.
An individual who has completed the programming exercises and derivations from Part 1 of Sutton and Barto's Reinforcement Learning book is seeking advanced practice problems. These suggested problems should be larger in scope than the textbook's toy examples but remain solvable using currently learned tabular methods. The goal is to gain more hands-on experience while continuing to master the theory before moving on to Part 2.
The author constructed a multi-agent prompt attention matrix manually to investigate data retention issues within complex language model interactions. The process involved significant mathematical computation, which the author described as "terrifying." This manual calculation was specifically designed to quantify how much information is lost during the intermediate steps of a multi-agent system. The core focus is on analyzing degradation or loss of context across multiple interacting agents powered by AI models.
Nanogate is a newly developed governance gate designed for AI agents. It operates with a specific runtime of 530 nanoseconds. The implementation language chosen for this system is Rust. This mechanism aims to govern the behavior and operations of autonomous AI agents efficiently. Details regarding its submission can be found via the provided link.
Due to record-breaking heat, air conditioning remains essential and projected to triple in number by 2050, according to the International Energy Agency. While AC units provide significant health benefits, such as preventing hundreds of thousands of premature deaths in 2019 alone, their increasing use poses an environmental problem. Scientists are considering newer technologies, like solid-state air conditioners, for a potential solution. The growing demand creates a conflict between human comfort and planetary sustainability.
Sundar Pichai is scheduled to deliver a commencement address at Stanford University in 2026. The speech will draw upon his experience and insights gained from leading major technology corporations. Attendees can read the details of his upcoming talk ahead of time. His message is expected to offer guidance and perspective to graduating students. Specific content or themes are not provided beyond the announcement of his address.
Scientists are developing advanced headphones designed to combat misophonia, a condition involving strong negative reactions to unwanted noises. This research, spearheaded by Shyam Gollakota at the University of Washington, utilizes machine learning technology for filtering sound. The goal is to create devices that can selectively eliminate specific triggering audio frequencies while preserving desirable sounds. For instance, users might block loud nearby conversations while still clearly hearing ambient sounds like birdsong. This represents a potential advancement in noise control for individuals sensitive to auditory irritants.
A new safety startup was launched by AI researchers due to concerns that "alignment is not on track." Sequent plans to manage a portfolio of underdeveloped research investments. The newsletter, Import AI, discusses various facets of AI research and relies on arXiv, cappuccinos, and reader feedback for its content. This issue specifically highlights the launch of this new safety initiative. Additionally, it mentions synthetic research interns as a topic of discussion.
To prevent Claude from being confidently incorrect, a guide advises including specific lines within a "Claude Skill." The original publication for this advice was Towards Data Science. The core recommendation focuses on four essential elements developers should embed into their skill programming. These required inclusions are necessary safeguards to ensure accurate and reliable output from the AI model.
World-Action Models (WAMs) represent an advancement in AI robotics, building upon pre-trained Vision-Language Model (VLM) backbones. A key component is the Vision-Language-Action (VLA) model, which functions as a robot policy generator. VLA models take knowledge from general VLMs and adapt them specifically for physical action execution by robots. This process allows systems to interpret complex language instructions alongside visual inputs to perform actions. Essentially, WAMs enable robots to bridge the gap between understanding (imagining) and doing (acting).
Addresses the computationally difficult open shop scheduling problem (OSSP) by proposing a novel Transformer-based scheduling policy utilizing an encoder-decoder architecture with multi-head attention. The model is trained on benchmark instances using only the processing-time matrix as input, achieving makespans near best-known solutions. When tested on significantly larger, randomly generated instances (40x40 to 100x100) without retraining, it showed an average relative gap of 12.89-15.12%. Compared to traditional dispatching heuristics like SPT and LPT, the Transformer proved substantially more effective. This demonstrates that a policy trained on smaller problems can effectively generalize to much larger OSSP instances.
Presents User Portrait based Nested Rollout Policy Adaptation (UP-NRPA), an online framework integrating Large Language Models for dialogue policy planning. This method overcomes limitations in current systems that cannot dynamically adjust to varied user traits like personality and objectives. Unlike prior methods requiring offline reinforcement learning per user group, UP-NRPA adapts strategies using real-time user feedback and a generated user portrait. Evaluations on collaborative and non-collaborative benchmarks showed significant gains, including a 56.41% increase in the sale-to-list ratio for negotiation tasks. Ultimately, UP-NRPA enables dialogue systems to adapt effectively to diverse users without retraining.
The Muddy Children Puzzle is a puzzle concerning knowledge and ignorance that significantly influenced the development of epistemic logic. Its original creator remains unknown, prompting an investigation tracing its roots through logical and literary publications spanning two centuries. The concept has inspired numerous variations, including adaptations involving numbers or colored hats. Furthermore, the study presents a novel version of a hats puzzle incorporating self-reference into the mechanics.
Agent swarms require sophisticated orchestration to handle complex tasks involving various data types. Existing frameworks struggle with the diverse input and interaction of multiple modalities like text, image, audio, and video (omnimodal scenarios). Orchestra-o1 is proposed as a new framework to enable efficient collaboration across these different modalities. It features a unified mechanism supporting modality-aware task decomposition, specialized sub-agents, and parallel execution. This design significantly improves performance on the OmniGAIA benchmark and utilizes DA-GRPO for state-of-the-art training.
Deep research tasks require agents to autonomously gather and integrate information in open environments, while agent evolution improves capabilities through environmental interaction but often lacks open-ended scope. The proposed Hybrid Open-Ended Tri-Evolution (HOTE) framework addresses this gap by using hybrid reinforcement learning for the collaborative evolution of three components: a proposer, a solver, and a judge. This process utilizes web-scale knowledge to facilitate development toward truly autonomous agents. Experiments on deep research benchmarks show that an 8B model trained with HOTE outperforms larger, static models from various advanced training methods. Furthermore, these results confirm that the joint evolution of all three modules within HOTE is crucial for success.
In a June 2026 revisit of the WorkBench benchmark, Claude Opus 4.8 emerged as the top agent, achieving an 89% task completion rate with only a 2.5% incidence of unintended harmful actions. This research indicates that for agents tested on WorkBench, improved capability correlates positively with increased safety performance. Despite overall progress and elimination of some error types, frontier models still exhibit basic mistakes capable of causing irreversible harm, such as erroneous emails. Furthermore, the availability of high-performing open-weight models has significantly reduced operational costs across the industry. Researchers have released an updated version of WorkBench incorporating new scores, data enhancements, and detailed analysis since 2024.
Researchers compared Difference-in-Means (DiM) interventions with two Iterative Nullspace Projection (INLP) techniques—nullspace projection and counterfactual flipping—on five chat models regarding safety refusals. INLP counterfactual flipping proved competitive with DiM directional ablation for suppressing refusal. However, nullspace projection was consistently weaker across the comparisons. Restricting INLP to leading directions maintains most suppression while keeping perplexity near baseline, showing potential tunability. Geometrically, the two INLP methods suggest different ways models encode concept absence versus its opposite.
YeasierAgent proposes an application-building paradigm centered on symbiotic agents, narrative worlds, and scene-aware interactions. It moves beyond traditional device-coupled software by defining applications as collaborative spaces involving users, agents, and worlds. The system facilitates rapid, cross-platform construction using platform-agnostic units like agents, scenes, and dialogue components. A key contribution is unifying intelligent agent capabilities—combining emotional companionship with practical tool execution—within one sandbox environment. This formalizes "Symbiotic Agent-Native Applications," shifting focus from isolated chatbots to cohesive, social computational environments.
TwinBI is an agentic digital-twin framework designed to improve multi-step analysis by unifying LLM interaction with executable BI dashboards. It addresses the issue of state inconsistency when users switch between direct dashboard manipulation and natural language queries. The system maintains a shared analytical state, tracking filters, hierarchies, metrics, and chart context via a unified interaction log. Evaluation shows TwinBI significantly increases exact-match accuracy from 43.3% to 63.3% and reduces timeout rates from 40.0% to 10.0% over a baseline. Findings indicate that this state-aware integration enhances both the reliability of the agent's analysis and overall user experience.
Recursive training on synthetic data risks model collapse by eroding distributional tails and homogenizing outputs. When verification relies on small, biased slices of data—common in proprietary silos like healthcare—the selection process itself becomes flawed. This siloed selection disproportionately keeps locally relevant samples while discarding globally important tail modes, thereby causing collapse rather than preventing it. Theoretically, this fragmented selection accelerates collapse and induces power-law diversity decay. A proposed initial mitigation involves creating Wasserstein proxy references from multiple silos without sharing the underlying raw data.
Gemma 4 models exhibit reproducible failure modes, looping on long factual enumeration tasks with high frequency. Researchers investigated whether these repetition loops could be corrected by editing specific model weights. By employing per-layer ablation and neuron attribution, the source was traced to small sets of MLP neurons or routed experts in larger models. Suppressing these identified neurons via static weight edits successfully mitigated the explicit looping patterns while maintaining general benchmark performance. However, the same edits failed to resolve more complex, fundamental "doom loops" involving an inability to recall specific facts, suggesting this failure is a knowledge limitation, not merely a circuit error.
Diffusion large language models (dLLMs) are appealing for low-latency mobile inference but face challenges efficiently utilizing Neural Processing Units (NPUs). The proposed llada.cpp is a novel NPU-aware framework designed to accelerate dLLM deployment on smartphones. It employs three key techniques: Multi-Block Speculative Decoding, Dual-Path Progressive Revision, and Swap-Optimized Memory Runtime. These methods adapt block-wise dLLM inference to match mobile NPU execution patterns and mitigate overheads from workload shrinkage and memory limitations. Testing demonstrates that llada.cpp significantly reduces LLaDA-8B generation latency by 17x to 42x compared to CPU baselines while maintaining quality.
Accurately predicting complex hypersonic flowfields, especially shock wave locations, poses significant challenges for current reduced-order models. A novel approach is introduced: a fully GPU-based workflow designed to address these limitations with high fidelity and low cost. This system combines accelerated data generation using a differentiable high-fidelity solver (JAX-Fluids) with neural emulator training. The method incorporates uncertainty quantification and uses physics-aware, residual-based refinement for improved physical consistency. These techniques result in robust emulators reliable even when extrapolating beyond their initial training conditions for practical engineering use.
Personalized federated learning addresses statistical heterogeneity by jointly training shared and personalized model parameters. A challenge in existing methods is that disparate local client objectives can lead to inconsistent updates for the shared parameters. To mitigate this, the paper proposes Federated Shared Parameter Correction (FedSPC), a modular correction technique. FedSPC applies control-variate correction exclusively to the shared parameters within various PFL architectures. Experiments on CIFAR-100 and Tiny-ImageNet demonstrate that FedSPC improves performance across multiple established PFL methods, such as FedPer and Ditto.
Grokking, the late generalization in neural networks, was investigated by intervening directly on the weight norm during training instead of merely observing its behavior. When trained with weight decay, networks exhibit grokking when the weight norm reaches a stable critical value ($W_c$). Holding the weight norm at a constant multiple ($\rho$) of $W_c$ still induces grokking, but the delay becomes proportional to $\exp(\alpha \rho)$. The derived exponential law suggests this pinned-norm delay is the counterpart to logarithmic delays seen with freely contracting norms. Finally, applying LayerNorm removes this explicit dependence by decoupling the weight scale from network functionality.
D2H-AD introduces a novel anomaly detection framework using Hyperdimensional Computing (HDC), addressing limitations of conventional methods like reliance on large labeled datasets. This system unifies distance-based similarity with density-aware encoding within the HDC paradigm for improved representation. Experimental results show D2H-AD outperforms five established baselines, including ODHD and Autoencoders, across multiple benchmark datasets. Its design is highly valuable due to being lightweight, interpretable, and computationally efficient through binary computations. These properties make it suitable for resource-constrained, real-time edge AI and TinyML applications.
Building upon low-rank adaptation (LoRA), this work proposes sparsity-induced extensions like Cheap LoRA (cLA) and ${c}^3$LA to enhance parameter efficiency. These methods introduce controlled structural restrictions by adapting within sparse, structured column subspaces. Researchers derived information-theoretic generalization error bounds for these advanced PEFT techniques. Empirical evaluation across 10 models and 14 datasets demonstrated that sparse LoRA adaptations remain competitive with their baselines. Crucially, this sparsity restriction achieves performance gains while simultaneously reducing training time by up to 10% and peak GPU memory by up to 15%.
Existing diffusion policy optimization methods suffer from instability because optimizing a variational surrogate can cause the Evidence Lower Bound (ELBO) to separate from the true log-likelihood, misaligning the resulting gradient. To address this "double-drift phenomenon," the authors propose DiPOD, a new framework designed for stable training. DiPOD maintains tight bounds throughout optimization by combining self-distillation with policy-improving gradient updates. This approach simplifies into augmenting every diffusion policy-gradient update with an on-policy ELBO regularizer. Empirical results show that DiPOD significantly stabilizes training and achieves higher rewards across both language model post-training and continuous-control tasks.
Cortical neurons show significant trial-to-trial variability, contrasting with highly consistent peripheral sensory neurons. While existing research suggests noise benefits discrimination in animals and machine learning, most ANN studies overlook correlations within the noise. Investigation reveals that structured, correlated noise significantly improves artificial neural network robustness against adversarial attacks and natural image modifications. Specifically, structure aids robustness to general image alterations, but this benefit is not universally transferable across different modification types. Overall, the findings suggest structured noise in ANN activations generally enhances robustness, providing a biologically plausible method for building reliable networks using only local information.
Introduces Localized Multidirectional Correction (LoMC), a framework designed for controlled post-training refusal suppression in routed Mixture-of-Experts (MoE) and hybrid-MoE foundation models. LoMC operates by first identifying a compact edit support, then aggregating prototype correction directions into layer-wise corrections. Finally, it applies rank-one, layer-wise correction exclusively within the specified support region. This gated approach enhances correction capacity without increasing the size of the intervention scope. Testing on various safety benchmarks confirms that LoMC significantly boosts non-refusal target-response behavior while preserving general model capabilities.
Recursive training of generative diffusion models can cause model collapse by drifting from the true data distribution. The research proves this recursion converges geometrically to a unique limiting distribution characterized as an infinite mixture of increasingly smoothed versions of the original data. Analysis shows that recursive training functions as a low-pass filter, significantly attenuating high-order spectral modes encoding fine structure. To counteract this compounding error, the authors propose annealed truncation schedules that progressively shrink truncation time across retraining rounds. These results are shown to be robust even when accounting for common discretization and score estimation errors during implementation.
Small variations in candidate tokens compound over thousands of steps during long-form language model reasoning. Existing fixed truncation methods fail to adapt to dynamic factors like entropy or task difficulty. Adaptive Nucleus Truncation Sampling (ANTS) addresses this by turning top-$n\sigma$ into an adaptive rollout-control mechanism. ANTS adjusts its selection window based on entropy and includes a safety fallback for training stability. On large-scale sparse MoE models, ANTS significantly boosts performance across various benchmarks, showing gains up to 10+ points on IFBench at long generation budgets.
Decision trees are crucial statistical tools known for their interpretability in classification and regression tasks. Bregman divergences offer a broad generalization of loss functions beyond simple Euclidean distance, encompassing measures like KL divergence. This work proposes extending the CART paradigm by utilizing these divergences to create a unified framework for decision tree construction. This approach allows various impurity criteria, traditionally model-specific, to be derived from shared convex and geometric principles. Furthermore, the authors investigate how properties of the generating convex function influence theoretical aspects such as impurity gains and estimator stability.
Optimal Transport is shown to be a powerful technique for domain adaptation by aligning source and target distributions. The study proves that using a $p$-norm cost with $p \geq 2$, the optimal transport map accurately recovers underlying linear transformations like rotations, translations, or homotheties in $\mathbb{R}^2$. A new method combining $K$-means clustering with optimal transport is developed to estimate this underlying map. This approach enables adapting linear regression models even when target data is scarce. The work emphasizes interpretability by applying these techniques to classical machine learning models rather than deep neural networks.
Addresses scenarios in scientific deployments with scarce labeled target data by leveraging pre-specified label-shift corrections from domain knowledge. Shows that the likelihood ratio between the corrected and source predictions forms a non-negative martingale, enabling an anytime-valid confirmation rule via Ville's inequality. This process transforms model monitoring into a formal sequential test based on the cumulative negative log-predictive density gap. Rejection indicates data conflict with the assumed correction relative to the original model, without quantifying the exact shift magnitude. Simulations confirm Type I control and power using Gaussian process regression for source models.
A novel double-space tensor-product RKHS framework is introduced to address quantifying hybrid aleatory and epistemic uncertainties in high-dimensional global sensitivity analysis (GSA). This method develops a concurrent double Möbius inversion to orthogonally decompose global dependence measures into pure aleatory effects, pure epistemic effects, and their interaction terms. The process uses an auxiliary-variable representation based on the inverse probability integral transform to handle hierarchical and Copula-induced correlations. The framework yields dimension-wise sensitivity indices that retain the structure of uncertainty attribution across all output dimensions. Numerical tests confirm the accuracy and scalability of this approach using examples like a multi-output Ishigami function.
Develops a statistical learning theory framework for applying gradient boosting to estimate covariate-dependent Generalized Pareto (GP) distributions for Peaks-over-Threshold modeling. By reparametrizing the GP likelihood and using Empirical Risk Minimization, non-asymptotic error bounds are derived accounting for statistical fluctuations, approximation bias, and finite iteration error. Simulations show this reparametrization improves convergence stability by reducing gradient correlation during training. The methodology was applied to a Texas Department of Insurance medical malpractice dataset with over 18,000 claims. Results indicate the number of days to settlement is the primary predictor influencing tail heaviness in settlement cost distributions.
Addresses the challenges of spatio-temporal forecasting involving irregular data, sparsity, and nonlocal interactions modeled by intractable integro-differential equations. Introduces the NonLocal Bayesian Spatio-Temporal model (NLBST), a hierarchical Bayesian framework for continuous fields learning explicit nonlocal coupling while maintaining tractability. NLBST utilizes a coordinate-based spatial basis expansion and models the coefficient process via an ODE derived from a Galerkin reduction of the nonlocal IDE, augmented by a Neural ODE residual for nonlinearity. It incorporates a linear-Gaussian observation model allowing Kalman-style updates under missing data and enables inductive prediction across unmeasured locations. Empirical tests confirm strong forecasting, spatial generalization, and well-calibrated uncertainty compared to baselines in challenging nonlocal regimes.
Evaluating predictive models before deployment is crucial for preventing algorithmic harm when decisions are informed by algorithms. Model performance can degrade due to covariate shift (changes in input data distribution) and selective labeling (outcome observability depending on historical decisions). The paper presents a double machine learning procedure to estimate a black-box model's target risk under both joint conditions. It identifies this estimand and derives a bias-corrected estimator using the influence function. Experiments on the eICU database confirm that this estimator more accurately tracks true target risk than methods addressing only one issue or combining standard approaches.