GLM-5.2 has been highlighted as performing well, passing a "vibe check" that suggests it's a true frontier open model. Multiple external validations reportedly rate GLM-5.2 highly, even above GPT 5.5 in one benchmark. This trend signals Z.ai's emergence as a significant and robust research lab in the field. The next major milestone anticipated is the release of an open Fable-class model from Chinese labs like Z.ai. Experts are speculating on when top laboratories might replicate such advanced releases given current industry constraints.
The primary focus in scaling frontier AI is shifting from acquiring more GPUs to maximizing utilization of existing hardware. Current models suggest best-in-class Model FLOPs Utilization (MFU) should reach 60–70%, contrasting with reports showing lower efficiency at major labs. The bottleneck has become a complex systems problem, involving scheduling, networking, and data pipelines, rather than solely capital expenditure. Anjney Midha discusses AMP’s vision for an independent compute grid designed to ensure efficient, reliable flow of computational power. Ultimately, achieving progress requires infrastructure that is more aligned, responsible, and functions like a regulated utility service.
The newsletter discusses recent developments in technology, business, and power. Key topics include Intel's alleged "free-market win" involving the White House supporting its chip development. It addresses global cybersecurity vulnerabilities, noting efforts like Amazon's announcement of AWS Continuum. Furthermore, it reports on Microsoft considering Chinese AI models and highlights a boiling point regarding data center backlash requiring new industry leadership. Overall, the content touches upon US government involvement with major tech companies and AI model access.
Scott Galloway, Ed Elson, and Jessica Tarlov have all posted new notes on Substack. Scott Galloway's latest note concerns Sweden settling an argument. Ed Elson shared notes regarding some significant numbers readers should see. Jessica Tarlov published a piece detailing "The UFC Moment That...". Readers can access these updates through the Substack app.
The decades-long hunt for dark matter, previously focused on WIMPs, has encountered difficulties from a "neutrino fog" created by cosmic neutrinos. Instead of stopping the search, this problem has prompted researchers to adopt new approaches. These updated methods include utilizing quantum sensors, liquid-helium detectors, and investigating signals within Jupiter's atmosphere. Additionally, entrepreneurs in Nairobi are adopting solar power for businesses because only about a quarter of Kenya's population lacks reliable centralized electricity.
Sonar earned recognition as a Leader in the Gartner Magic Quadrant for Technical Debt Management for 2026, achieving the highest rating on Ability to Execute. Unlike reactive tools, SonarQube focuses on preventing technical debt before it enters the codebase and automatically remediating issues within the development loop using its AC/DC framework. The platform also analyzes code reliability via its LLM code quality leaderboard and embeds context into custom models through SonarSweep. It offers extensive enterprise capabilities, supporting over 40 languages and integrating across the entire Software Development Life Cycle (SDLC). Currently, the solution is relied upon by millions of developers and a significant portion of Fortune 100 companies.
Physicists' long-standing hunt for WIMPs, a leading dark matter candidate, is being reinvigorated. The latest edition of "The Download" newsletter reports on advancements in this scientific pursuit. In addition to the particle physics focus, the publication also addresses energy initiatives. Specifically, it features coverage regarding Kenya's strong push toward adopting solar power. This combination highlights both cutting-edge science and renewable energy developments.
OpenAI has released enhanced features for ChatGPT Enterprise to aid organizational management of AI resources. These additions include comprehensive usage analytics, providing detailed insights into how the platform is being utilized. Furthermore, new spend controls are available, giving enterprises direct tools to manage and monitor associated costs. This combination allows organizations to scale their AI adoption more securely. Ultimately, these updates help businesses manage expenses while confidently expanding their use of ChatGPT Enterprise.
FERC issued a significant ruling concerning large-load interconnections for new energy consumers. This action directly impacts how facilities such as AI factories, semiconductor fabs, and advanced manufacturing sites can connect to the electrical grid. The decision addresses the increasing strain placed on infrastructure by burgeoning technologies like artificial intelligence. Implementing these guidelines aims to improve overall grid reliability and enhance energy affordability. Such updates are crucial for supporting continued technological growth within the power sector.
Governmental and corporate focus is increasingly shifting toward achieving "AI sovereignty." This trend emphasizes the need for greater control and secure access regarding artificial intelligence technologies. Anthropic's current model restrictions exemplify this rising concern. Entities are actively prioritizing self-sufficiency in their AI infrastructure. This signals a maturation phase in the ongoing race to develop advanced AI systems.
Google Cloud heavily emphasizes agentic AI within its overall enterprise strategy. This focus indicates a significant commitment to integrating autonomous, goal-oriented AI agents into business solutions. Such systems are designed to perform complex tasks with minimal human intervention. The betting suggests these advanced AI capabilities will be central to future cloud offerings for businesses. Ultimately, the platform aims to power an "agentic enterprise."
Stellantis, Wayve, and Uber are partnering to develop a global robotaxi service. This collaboration combines expertise from major players in the automotive and autonomous technology sectors. The venture aims to establish a worldwide network of self-driving ride-hailing services. By uniting their resources, the partners plan to advance Level 4 autonomy on a massive scale.
Adobe has embedded "creative agents" across Creative Cloud applications like Photoshop and Premiere Pro, moving beyond simple media generation. This agent acts as an orchestration layer by executing complex, multi-step production workflows using deep API access. New features include "Elements" for visual consistency and "Projects" for managing contextual memory across sessions. The AI automates tedious tasks such as batch renaming video clips or applying brand updates across multiple layouts. Adobe positions this technology to enhance human designers' roles as "creative directors," while integrating the agent into third-party platforms like ChatGPT.
Jennifer Siebel Newsom promoted her new film, *Miss Representation: Rise Up*, which examines the backlash faced by women due to algorithms and deepfakes. During an interview near the White House, her husband, Governor Gavin Newsom, announced that the couple was under a Justice Department investigation. This investigation specifically targets Siebel Newsom's taxes and the California Partners Project, a non-profit she co-founded. The project received $4.3 million in donations solicited by Gavin Newsom. Governor Newsom dismissed the investigation as a "personal vendetta" potentially linked to Donald Trump concerning his own presidential aspirations.
Marc Isaacs' film *Synthetic Sincerity* explores the potential overreach of artificial intelligence using a fictional premise of an AI lab harvesting human emotion from cinema. The research licensed Isaacs' entire documentary body, including works about London life (*Lift*) and the town of Frinton-on-Sea. Data analysts at the University of Southern England were permitted to feed these documentaries into their system for character creation. Isaacs also incorporated the chef/owner of his lunch venue, Ablikim Rahman, into the film as a subject for AI capture. Ultimately, Isaacs admits the university itself was fictionalized for the project.
Version 0.6a0 of datasette-acl expands its functionality beyond just table-only permissions. It evolves into a more comprehensive resource-sharing system for Datasette instances. A key addition is a new plugin that enables granular control over resource access. This allows administrators to manage user permissions across various resources within the multi-user Datasette environment. Alex Garcia was primarily responsible for developing these new features.
Miami-based AI startup Subquadratic recently revealed itself with a significant claim. The company announced it solved a mathematical bottleneck believed to have constrained large language models for nearly ten years. Initial details provided by the startup were sparse, leading to some skepticism from the public. However, Subquadratic has since begun providing supporting evidence regarding its breakthrough.
Introducing an AI agent named Ask Ad Manager. This tool is positioned within the context of Agentic AI for Google Ad Manager. Visually, its introduction is marked by an animated sequence on a white background. The branding features a blue magnifying glass paired with a spark icon. Ultimately, it offers assistance to help users achieve greater productivity with ad management tasks.
NVIDIA showcased how AI is driving a shift toward autonomous operations in advertising and marketing. Building advanced technology for this sector requires infrastructure capable of handling immense speed and scale. The necessity of adopting AI is now understood across the industry. A key focus at Cannes Lions was evaluating whether existing infrastructures could support these demanding new requirements. This highlights a critical technological pivot point for the future of the creative industries.
Anthropic's Claude Code Artifacts feature enables users on Team/Enterprise plans to generate live, interactive, and shareable HTML webpages directly from their code sessions. These artifacts dynamically surface the results of code execution and connected data sources in real-time at a single URL. This capability bridges technical engineering work with non-technical stakeholders by eliminating the need for manual status updates or external infrastructure setup. While structurally limited to self-contained, ephemeral captures, this contrasts with OpenAI's Codex Sites, which builds persistent, full-stack web applications. Overall, both platforms aim to revolutionize how AI outputs are communicated in an enterprise setting through secure, shared visual workspaces.
Introducing Arbor, a new framework designed to transform AI optimization from trial-and-error guesswork into cumulative learning. It achieves this by organizing research hypotheses, experiments, and insights within a persistent, branching "Hypothesis Tree." This structure separates the strategic planning (Coordinator) from isolated coding tests (Executors), allowing simultaneous testing of multiple competing ideas. In practical evaluations, Arbor surpassed standard AI coding agents by over 2.5 times on real-world engineering tasks under the same compute budget. By enforcing an explicit merge gate against held-out evaluators, Arbor ensures demonstrable, transferable progress rather than mere development overfitting.
Liquid AI has launched two new models: LFM2.5-Embedding-350M and LFM2.5-ColBERT-350M. These retrievers integrate both a dense bi-encoder and a ColBERT late-interaction model. Their combined architecture is designed for efficient multilingual search functionality. The system supports searches across 11 different languages. A key advantage of these models is their suitability for deployment on edge devices, ensuring fast performance.
Perplexity introduced Brain, a new self-improving memory system for its Computer agent. This system focuses on remembering the agent's own work history rather than just user interactions. Brain constructs a traceable context graph detailing successful outcomes, failures, and necessary corrections. It reviews this accumulated data overnight to enhance performance. Initial results show improvements in the agent’s correctness, recall, and cost efficiency.
Google Gemini can save significant time during meeting preparation. This functionality assists users in streamlining their meeting workflow. The described capability integrates AI assistance directly into the preparation process. Utilizing Gemini allows individuals to spend less time and more efficiently prepare for meetings.
PCIe transfer latency is identified as a bottleneck in agentic inference workflows. A custom, device-resident vector search kernel was developed to address this issue. Building this kernel bypasses the CPU during the retrieval step. This optimization unlocks deterministic microsecond tail latencies for the process. The work demonstrates how to improve performance critical steps in Retrieval Augmented Generation (RAG) agents.
Reliably extracting structured data from Large Language Models (LLMs) is crucial for application integration. Two primary methods discussed are using dedicated JSON Mode and implementing Function Calling. JSON Mode forces the model output into a specific JSON format, ensuring structural consistency. Function Calling allows models to generate explicit function calls with necessary parameters, enabling programmatic interaction with external tools. Choosing between these two depends on whether the goal is structured data extraction (JSON Mode) or triggering defined actions (Function Calling).
Examines the strengths and weaknesses of Claude Fable 5 specifically concerning coding abilities. The content provides an analysis to help users understand its capabilities in programming tasks. It was initially published on the Toward Data Science platform. The focus is a detailed evaluation of the model's performance for developers.
Traditionally, proteins were understood to feature a hydrophobic core housing nonpolar amino acids. New research suggests an expansion of this structural understanding beyond just the hydrophobic core. Specifically, other types of amino acids—such as polar, acidic, and basic residues—also exhibit clustering behavior. This pattern describes these different chemical groups aggregating together in approximate units of eight. Researchers have developed methods to quantify and visualize this newly identified "Mosaic Q model" within protein structures.
Discusses the processes involved in dispatching parsed Retrieval Augmented Generation (RAG) questions using document profiles. It details three approaches for deciding which functions or logic "fire" based on the input. The content also covers the concept of activations and an associated audit meta block. Additionally, it provides a walkthrough involving the broker and corpus. This information is presented in Enterprise Document Intelligence, Volume 1, Issue 6c.
Provides a guide for implementing image similarity search using Milvus. It discusses the limitations of simple visual matching when performing such searches. The content specifically addresses the practical aspects of setting up vector-based image retrieval systems. Understanding these technical nuances is crucial because relying solely on visible similarities can be insufficient for accurate results.
An undergraduate AI student is seeking guidance on conducting reinforcement learning or world model research. The core confusion revolves around the practical process of how to actually conduct these experiments. The post explicitly asks for advice regarding what genuine research entails in this field. No specific methods are outlined, only a request for foundational knowledge. Advice from experienced individuals is requested to guide their research efforts.
Employing sound waves allows neuromorphic devices to better mimic biological neurons and achieve superior energy efficiency over existing electronic counterparts. Scientists developed an acoustic synapse using phase bits ($\phi$-bits) that support parallel computing and multiple simultaneous computations within a single space. This novel device demonstrated improved performance, achieving high accuracy in pattern recognition with fewer parameters and faster training times than conventional chips. The technology successfully simulated synaptic plasticity and incorporated neuromodulatory effects by adding extra rods to the acoustic structure. Overall, this acoustic approach could lead to future hardware consuming significantly less power for complex tasks like sensory processing and data analysis.
The primary focus of modern AI scaling should shift from merely acquiring more GPUs to maximizing existing computational power. Poor utilization rates, such as xAI reportedly running below 10% Model FLOPs Utilization (MFU), indicate that the issue is systemic rather than purely hardware-related. Frontier AI advancement is increasingly a complex systems problem involving scheduling, networking, and data pipelines, not just capital expenditure on GPUs. The discussion explores concepts like "output maxxing" and AMP's compute grid vision to create more efficient infrastructure. Ultimately, the industry needs better alignment across the entire stack for sustainable progress in artificial intelligence.
Anjney Midha's LinkedIn profile indicates she is located in San Francisco, California, and lists AMP PBC and Stanford University under experience and education sections. The visible information suggests a professional network with over 26K followers and 500+ connections. Accessing the full details, including specific titles and tenure, requires signing into LinkedIn. The platform also offers functionality for connection management, such as viewing mutual contacts and requesting introductions from Anjney at AMP PBC.
Metrics offer substantial insights into various aspects of life. However, they possess a dual nature, capable of revealing crucial information while simultaneously obscuring or corrupting understanding. The author spent over ten years meticulously tracking personal details to fully grasp this inherent duality. This experience suggests that the very act of measurement has implications for what is perceived as reality. Ultimately, this highlights both something about detailed self-tracking and the fundamental nature of quantification itself.
Casey Harrell, a man with ALS, is noted by researchers as the "first power user" of a brain implant. This technology allows him to communicate despite being paralyzed and unable to speak clearly otherwise. For nearly three years, Harrell has been using a brain-computer interface (BCI). The development highlights significant advancements in neural technology for individuals with severe motor impairments.
Museum debuting an immersive "Infinity Room" powered by artificial intelligence. This exhibit showcases art created entirely through AI generation. Visitors will experience digital artworks that adapt within a vast, endless visual space. The integration of AI technology revolutionizes traditional gallery viewing experiences. Such exhibits represent a major shift in how artistic media is displayed and consumed.
Describes an end-to-end workflow for Salesforce CodeGen, utilizing a model loaded from Hugging Face. The process extends beyond simple inference by incorporating function extraction, syntax checking, static safety validations, and unit-test execution. Further enhancements include reranking the top N generated candidates and synthesizing multi-turn programs. Researchers also experiment with different prompt styles to improve generation quality. Finally, the workflow culminates in visualizing a mini benchmark and exporting the resulting artifacts for reuse.
Excessive reliance on chatbots has been found by MIT researchers to potentially diminish critical-thinking skills. Furthermore, such dependence could reduce an individual's innate ability to discern misinformation. With increasingly common manipulated images and misleading headlines due to advanced AI, the danger is heightened. Over-relying on artificial intelligence for verifying online content can compromise personal judgment skills. While AI offers tools to detect fake material, the research highlights a potential drawback to its over-dependence.
Datasette Apps are self-contained HTML/JavaScript applications running within a restricted `<iframe sandbox>` inside Datasette. These apps can execute read-only SQL queries against the underlying data, and optionally perform writes using pre-configured stored queries. The implementation uses advanced security measures like `sandbox` attributes and Content Security Policy (CSP) headers to prevent data exfiltration or unauthorized access. Communication is managed securely via MessageChannel() for controlled interaction with the parent Datasette application. Developers can generate boilerplate code prompts within the plugin, facilitating creation by Large Language Models.
MosaicLeaks investigates the potential security vulnerabilities of automated research agents handling sensitive information. It tests whether these AI tools maintain confidentiality when processing private data. The methods reveal that current systems can inadvertently leak confidential material to unauthorized parties or model outputs. Researchers found specific scenarios where prompts and system behaviors led to unintentional disclosures. Consequently, implementing stronger privacy controls and verification protocols is crucial for deploying such advanced research agents safely.
Autonomous agentic AI systems pose novel security and compliance risks due to their capabilities, such as tool invocation and cross-agent coordination. Governing these agents requires more than simple access control, necessitating rules defining permissions, obligations following actions, and conflict resolution mechanisms. Existing policy engines like XACML or Rego are insufficient because they lack support for obligation lifecycle management, meta-policy conflict resolution, or waivers. Proposes AgenticRei, a system utilizing a deontic policy language built on OWL to enforce these advanced governance constraints at runtime. This approach evaluates policies via an external logic engine, governing both agent tool use and inter-agent communication while composing with existing industry frameworks.
Developed a human-in-the-loop pipeline to longitudinally measure how well an undergraduate computer science program covers external knowledge guidelines across revisions (CS2013 and CS2023). This pipeline processes structured corpora for semantic retrieval, followed by human validation of course-to-knowledge matches. The study found the program covered approximately 50% of knowledge units under both curricula over a decade. While competency was articulated for about 88% of covered units, cognitive depth coverage lagged, showing lower rates under CS2023 compared to CS2013. The framework successfully separated persistent structural curriculum gaps from natural evolution differences between the standards.
Diffusion Language Models (DLMs) offer a text generation alternative to autoregressive LLMs by using iterative denoising for parallel sequence refinement. A systematic experimental analysis evaluates eight state-of-the-art DLMs across eight benchmarks covering reasoning, coding, translation, and more. The study specifically assesses both generation quality and computational efficiency for these models. Furthermore, it analyzes the influence of key inference factors like denoising steps and context length on performance. Findings indicate that a DLM's behavior is heavily dependent on generation-time design choices, creating distinct trade-offs between performance and required computation.
Multi-agent LLM deliberation, used to boost reasoning, is modeled as a closed-loop dynamical system incorporating agents' hidden internal beliefs called anchors. These anchors exert a continuous pull on an agent's opinion independent of neighboring input. Researchers demonstrate that this individual anchor can be recovered solely from the observed deliberation process. This mechanism explains why agents' confidence can exceed the initial space (convex hull) formed by all starting opinions. The recoverability and predictive power of the anchor provide a useful test for when true internal belief drives the system beyond simple consensus rules.
DeXposure-Claw is introduced as an agentic system designed for supervising fast-moving, networked credit risks in decentralized finance (DeFi). It addresses limitations of general-purpose LLM agents by routing their decisions through structured evidence for better risk assessment. The system integrates DeXposure-FM, a graph time-series foundation model, to forecast future exposure networks. Deterministic monitors and stress scenarios then convert these forecasts into typed alerts, attribution signals, and scenario evidence. Finally, data-health and confidence gates constrain escalation before generating auditable supervisory tickets, which are validated using the DeXposure-Bench evaluation harness against ground truth losses.
Investigating LLM knowledge limits on clinical tabular data using cross-model attribution divergence, the study compared Qwen 2.5 7B with XGBoost. Verbalized LLM confidence proved unreliable, remaining high irrespective of actual predictive accuracy. The models showed an inverse difficulty effect, performing worse when XGBoost was highly accurate but matching it under moderate uncertainty. Adding few-shot examples and SHAP evidence significantly reduced disagreement scores and improved accuracy without retraining. A novel cross-model calibrator using attribution divergence successfully replaced uninformative verbal confidence with accurate patient-specific reliability estimates.
Retinal imaging presents a noninvasive means to assess Alzheimer's disease (AD) risk using vision-language frameworks. Existing methods treating phenotypic similarity via hard group assignments limit representation learning. This research proposes a continuous formulation for phenotypic structure within contrastive learning, replacing rigid grouping with differentiable weighting functions based on intra-modality embeddings. These weights create soft multi-positive relationships, offering graded supervision for disease risk assessment. Evaluations show this continuous approach outperforms discrete grouping methods when predicting incident AD using retinal and clinical data from the UK Biobank.
Researchers developed an online technique called "conscience step" to enable Large Language Models (LLMs) to review and correct their own unethical reasoning. This involves extending the training loss with an alignment component using Direct Preference Optimization (DPO) to guide the model toward ethical outputs. The method is versatile, applicable across various settings including fine-tuning, adversarial prompting, and zero-shot learning. Importantly, it functions without requiring a separate judge by referencing a frozen copy of itself. Instead of observing emergent misalignments like code hacking, the study demonstrates how single introspective questions can steer training toward verifiable "Emergent Alignment."
Introduces ITNet, an Integral Transform Network built around a learnable kernel jointly depending on positions and features. This unified architecture shows that established methods like convolution, self-attention, and recurrence are special cases of this underlying mathematical transform. The MLP structure allows the model to adapt its pairwise interactions directly from data. Developed techniques include tiled kernel fusion and importance-weighted Monte Carlo integration for efficient computation. ITNet achieves performance matching or exceeding specialized models across diverse tasks including ImageNet and GLUE.
Traditional uncertainty frameworks are insufficient for interactive LLM agents, necessitating new methods for proactive clarification seeking. Given practical constraints like black-box APIs, prompt-based estimation emerges as the most viable technique for signaling uncertainty at deployment time. The paper introduces a simple prompt-based decomposition that specifically separates action confidence from request ambiguity. Evaluation on two custom benchmarks showed significant improvements: the proposed method increased clarification F1 on ALFWorld-Clarification by 73% over ReAct+UE and 36% over UAM, generalizing across five different LLM backbones. These results demonstrate the efficacy of prompt-based decomposition for enhancing agents' ability to handle underspecified tasks.
Identification conditions quantify whether a target query or parameter can be computed from available information, often represented by causal graphs and partial data. Theoretical identifiability relies on idealized assumptions like infinite data to define the unique determination of causal effects. This paper introduces "computational identifiability" as an alternative framework that defines a finite computational search procedure for empirical estimators. Identifiability is satisfied if this procedure finds an estimator within a specified error tolerance, contingent on the chosen search assumptions and procedure. Experiments show this method effectively answers practical identification questions involving small samples, mixed data, and ambiguous criteria.
Deployment of large Time-Series Foundation Models in science is limited by distributional misalignment and computational cost. The paper introduces Gated Uncertainty-Aware Routing for Distillation (Guard), a framework to extract knowledge from misaligned models. Guard treats multi-teacher distillation as an instance-wise decision process using two novel components: a Contextual Router and an Uncertainty-Gated Temperature mechanism. These mechanisms adaptively select the best teacher and temper distillation strength when teacher confidence mismatches domain reality. Evaluation across four climate domains shows Guard significantly reduces forecasting error compared to fixed-weight baselines, enabling edge deployment.
Social scaffolding in user prompts creates a "Social-Semantic Gap" during Large Language Model (LLM) inference, contributing to high energy costs. SPSD (Sentiment Preserving Semantic Distillation) is an edge-based pipeline that compresses these prompts using a quantized Small Language Model prior to cloud transmission. Testing on 248 prompts showed an average saving of 99.9 input tokens per call while maintaining response quality within a one-point margin across a 15-point rubric. This approach can reduce the energy cost for LLM inference by an estimated 70–270 uWh per call under assumed conditions. Consequently, SPSD demonstrates that on-device distillation effectively reduces cloud LLM input token costs without degrading response quality significantly.
State-of-the-art medical diffusion models are resource-intensive due to complex computations across many U-Net evaluations. Performance analysis of Med-DDPM revealed that training is bottlenecked by cuDNN convolution and implicit-GEMM kernels due to memory access patterns and suboptimal Tensor Core usage. Two proposed architecture-aware optimizations—TF32 Tensor Core activation and a 3D channels-last layout—were tested against different NVIDIA GPUs. These optimizations significantly improved performance metrics, reducing SM cycles and dynamic instructions by up to 100x. Furthermore, the changes substantially increased Tensor Core utilization and overall IPC on A100 without impacting synthesis quality.
Information Lattice Learning (ILL) interprets a signal by projecting it onto a partition lattice encoding abstraction hierarchies, then lifting learned rules back to the original domain. When applied to probability mass functions, the probabilistic rules derived from ILL naturally correspond to constraints in a Probabilistic Graphical Model (PGM). These rule sets represent marginal constraints over interpretable abstractions defined by the lattice's partitions. The feasible family of joint distributions satisfying these constraints is addressed via "general lifting," often implemented using an L2 uniformity principle related to maximum entropy. Therefore, ILL is best viewed as structure learning for constraint-based factor graphs built upon quotient variables, rather than a standard Bayesian network.
A study examines the evolution of the Weibull weight-scale parameter ($\lambda$) during AdamW training using a two-parameter Weibull framework. The research decomposes the squared weight norm's change into three forces: alignment, injection, and decay. On Pythia-70M models, the alignment force dominates the initial rise phase, contributing 88-94% of the total absolute force budget. At near saturation, the balance between alignment and decay explains the transition to parameter relaxation. The authors propose a spline displacement method to estimate the crucial alignment force even when optimizer moments are unavailable.
Estimation-of-distribution algorithms (EDAs) typically struggle when applied to sparse parameter spaces because existing methods require handcrafted sparsity operators. The authors propose using multivariate zero-inflated Gaussian (ZIG) distributions as the sampling law for EDAs, enabling optimization in sparse settings without such restrictive assumptions. ZIG models account for both sparsity patterns and active parameter values simultaneously through a latent Gaussian structure. They introduce practical amortized inversion-based estimators that accurately recover these latent correlation structures. Testing on the Lunar Lander benchmark showed that ZIG-EDA converges faster and achieves superior results compared to dense or existing sparse evolutionary algorithms.
Self-play reinforcement learning offers a path to training driving policies using only simulations, bypassing the need for extensive human data. A limitation of pure self-play is that learned behaviors can become incompatible with actual human driving conventions. This method improves upon existing approaches by treating limited human demonstrations as a regularization objective alongside a minimal safe goal reward. Remarkably, it achieves alignment using just 30 minutes of human data, significantly less than comparable imitation learning methods. Trained policies demonstrate coordination with unseen human trajectories efficiently on consumer-grade hardware.
Alzheimer's disease diagnosis increasingly relies on expensive multimodal data like structural MRI and PET scans. ProMUSE introduces a Progressive Multi-modal Uncertainty Guided Staged Evidential Network to address the impracticality of full-modality inference in clinics. Initially, it classifies using low-cost clinical data while quantifying uncertainty with a Dirichlet-based model. If uncertainty is high, ProMUSE adaptively incorporates MRI or PET features, fusing beliefs via Dempster-Shafer theory for a calibrated prediction. Testing across multiple AD datasets confirmed ProMUSE achieves competitive accuracy while significantly reducing the need for costly imaging by 50-90%.
Ventricular tachycardia ablation requires precise pace-mapping, which is manually intensive and time-consuming. Existing AI methods for guidance fail to transfer knowledge across different ventricular tachycardia (VT) targets within the same patient. cAPM introduces a novel framework employing continual learning and a task-agnostic surrogate neural network to accumulate and utilize data from previous mapping sessions. This allows it to efficiently guide clinicians by identifying the most informative pacing sites sequentially. In simulations, cAPM significantly outperformed state-of-the-art methods, localizing targets with 81% accuracy using only 4.5 pace-mapping sites compared to 38% using 13.7 sites.
Signed pairwise interaction scores incorrectly conflate uniqueness, redundancy, and synergy. A structural causal model demonstrating this failure shows faithful indices returning zero per pair, unlike projective indices which combine all effects into pair scalars. The authors propose Stochastic Hi-Fi, a retraining-free method using interventional masked inference to estimate individual feature U/R/S profiles. This estimator offers exact interventional semantics and proven statistical properties. Empirical testing demonstrates Stochastic Hi-Fi significantly outperforms scalar methods on tabular SCMs and successfully separates interaction mechanisms in models like GPT-2.
Presents a scalable, solver-free training method for models operating in the predict-then-optimize paradigm. This approach addresses the intractability of directly minimizing empirical decision regret when using linear or combinatorial optimization. Unlike existing methods that require computationally expensive solver calls during gradient evaluation, this technique uses a measure transformation principle. The result is a novel surrogate loss function that requires no iterative optimization solving during training. Theoretical guarantees for the method include Fisher consistency and excess risk bounds, with empirical results showing superior speed and performance.
Presents a comprehensive pipeline for Bayesian mixture model inference in federated learning settings where data cannot be fully shared. It utilizes an extended Variational Consensus Monte Carlo (CMC) approach to estimate local posteriors independently across data silos and aggregate them. Key contributions include adapting variational CMC for mixtures that infer both the number of clusters and all parameters, even without assuming conjugacy. The method also introduces novel cluster-matching algorithms for cross-silo settings where not all clusters are present locally. A simulation study validates this framework, showing superior recovery of small clusters compared to standard pooled MCMC methods when local data reflects underlying structure, using electronic health record data as an example.
AURA introduces an adaptive framework to audit LLMs used as judges for open-ended generation tasks. It addresses the limitation of existing pipelines that require reliable, pre-existing supervision signals, which are often unavailable in practice. AURA operates on pairwise LLM-as-a-judge decisions, utilizing limited human verification data. The core mechanism is iteratively learning a "human-consistency signal" to progressively refine trust in the judge model. This allows the system to prioritize uncertain comparisons for targeted human review, ensuring robust evaluation with scarce supervision.
Addresses stochastic linear contextual bandits (SLCB) under the specific assumption of bounded reward noise. Existing literature often assumes weaker sub-Gaussian noise, yielding $\tilde{O}(\sqrt{T})$ regret bounds. The paper proposes a novel algorithm named SME-OFU that integrates set-membership estimation with optimism in the face of uncertainty. This new approach achieves an improved regret bound of $O(\log T)$. Empirical simulations validate that SME-OFU outperforms benchmark algorithms when reward noise is bounded.
A novel active-learning framework combines Active Kriging Monte Carlo simulation with conformal prediction for structural reliability failure probability estimation. The approach uses an adaptive cross-conformal strategy optimized for small samples, employing a J+GP conformal estimator. This method provides distribution-free guarantees on prediction errors, surpassing standard AK-MCS techniques. Such improvements enhance the accuracy and robustness of failure estimates, particularly in rare-event regimes near the limit-state surface. Numerical comparisons confirm the effectiveness of this enhanced uncertainty quantification framework.
Investigates off-policy evaluation (OPE) in Markov decision processes when immediate rewards are Missing Not at Random (MNAR), which violates ignorability and causes selection bias. To address this, the work formalizes a reward-dependent propensity model utilizing future states as shadow variables to identify the full-data conditional mean reward. A novel bridge function is introduced to recover this conditional mean reward via a min-max procedure, avoiding explicit MNAR modeling. An accompanying Fitted-Q-Evaluation-style estimator propagates these recovered rewards and allows target policies to incorporate missingness indicators. The developed estimator demonstrates consistency and finite-sample error bounds, showing strong performance on simulated and real sepsis data.
Deep learning has achieved superior real-world performance despite contradicting classical statistical intuitions. This paper examines key features and unexpected aspects of deep learning using a physics-informed viewpoint. It carefully identifies and justifies the various choices involved in constructing deep learning models. Specifically, it reviews neural scaling laws. The discussion further explores how these scaling laws interact with constraints and inductive biases when applying machine learning to physical problems.
SSH-Net, a Structured Segmented Hazard Deep Neural Network, is proposed for predicting failure time distributions under competing risks. It addresses limitations in existing deep learning models by structuring the network architecture to match complex physical system data. The model utilizes separate sub-networks allowing different covariate groups to independently influence failure prediction within a cause-specific framework. SSH-Net outputs cause-specific hazard functions and minimizes a penalized log-likelihood loss function. Its performance was validated using simulation studies and applied successfully to predict GPU failure time distributions using Titan data.