Research updates reveal the next wave of product capabilities and infrastructure needs. This hub follows papers, benchmarks, datasets, lab systems, releases, and open reproductions, focusing on which results may reach model training, agent systems, robotics, or developer tools.
The author, a developer productivity expert, examines multiple studies showing that AI tools make developers feel more productive but don't actually speed up shipping. Bottlenecks have shifted to code review, CI/CD, QA, and other downstream processes. The article offers recommendations: stricter code review, CI adaptation, feature flagging, protecting knowledge sharing, and more.
METR's controlled experiment found AI-assisted developers took 19% longer, yet felt 20% faster.
A follow-up study collapsed because many developers refused to work without AI, creating selection bias.
Over 200 Nobel laureates and global leaders gathered in Rome for a three-day assembly on AI and nuclear war, culminating in the signing of the Rome Declaration emphasizing human dignity, cooperation, and peace. The event, inspired by Pope Leo XIV's encyclical, featured addresses by prominent figures including Juan Manuel Santos and Muhammad Yunus, focusing on the urgent need for AI governance.
The Global Nobel Laureates Assembly convened 200+ Nobel winners to address AI and nuclear war risks.
Inspired by Pope Leo XIV's encyclical Magnifica Humanitas.
Free tool that scans your website to score its readiness for AI agents and LLMs across three tiers: Access, Citation, and Transaction, with actionable fixes.
Free, no account required
Evaluates Access, Citation, and Transaction readiness
Neko Health raised $700 million in Series C funding to launch its AI-powered preventive health screening service in the United States, starting with a New York clinic. The company combines full-body scans, blood tests, and clinician review.
Neko Health raised $700 million to expand its AI body scan service to the US.
The funding round was led by Lightspeed and O.G. Venture Partners, with participation from celebrities.
Glad-AI-Tor is a platform that ranks AI tools based on crowd verdicts. It covers voice, LLMs, image, video, coding, and music with 75 tools and 188 votes. Rankings are unbiased and cannot be bought.
Six categories: voice, LLMs, image, video, coding, music.
Rankings based on real visitor verdicts, one vote per tool per person.
VaultCharts is a free desktop trading app that combines charting tools with an AI assistant. It supports multiple AI models, is local-first, and allows users to analyze markets with or without AI assistance.
VaultCharts offers a free desktop trading app with charting tools and an AI assistant.
Users can bring their own AI model or use local models like Ollama.
A new minimum-lap-time planning framework incorporates robustness to state disturbances and parameter uncertainty, validated on simulated FSAE car using MPC.
Extends prior disturbance-aware MLTP to account for uncertainty in moment of inertia, center-of-mass position, and aerodynamic drag.
Uses a parsimonious activation strategy to apply robust constraints only where critical, maintaining computational tractability.
This paper proposes a barometer-aided attitude estimation architecture that uses barometric altitude measurements to provide complementary vertical motion information, enhancing attitude estimation within nonlinear observers on SO(3). Two observers are designed: a deterministic Riccati observer cascaded with a complementary filter achieving almost-global asymptotic stability, and a nonlinear observer on SO(3)×R2 guaranteeing local exponential stability. Simulations and real flight data validate the method's effectiveness under minimal-sensing configurations.
Introduces barometer-aided attitude estimation using barometric altitude for vertical motion information.
Designs a deterministic Riccati observer cascaded with complementary filter for almost-global asymptotic stability.
WANDA is a synthetic data engine that learns open-world mobile manipulation policies from a single demonstration. It reconstructs background and interaction trajectories, rearranges configurations, uses Corrective State Expansion for robustness, and synthesizes trajectories on diverse 3D worlds, achieving generalization and cross-embodiment support.
WANDA generates abundant synthetic training data from just one real demonstration.
Utilizes Gaussian splatting and whole-body motion planning for data synthesis.
This paper introduces HRO, a hierarchical room-to-object framework for zero-shot object-goal navigation powered by large language models (LLMs). Unlike existing flat reasoning methods, HRO mimics human-like hierarchical spatial cognition, enabling the agent to explore from room-level to object-level in a coarse-to-fine manner. Experiments on Gibson and HM3D datasets demonstrate superior success rate and generalization over prior LLM-based approaches.
A multimodal teleoperation architecture for ROVs using 3D Gaussian Splatting to generate occlusion-free exocentric views and a vibrotactile suit for haptic cues. A human study with 30 participants showed the exocentric view significantly improves performance under high latency, with fNIRS indicating sustained executive control rather than cognitive overload.
DAVS uses real-time 3D Gaussian Splatting to create an occlusion-free exocentric viewpoint
Researchers propose a hierarchical Bayesian generative model that operationalizes uncanny valley guidelines as mathematical design variables. The model maps the effect onto four variables: deviation from predicted robot-category mean, inconsistency in human likeness across modalities, prediction uncertainty, and observational uncertainty. Experiments show that increased observational uncertainty attenuates familiarity dips at intermediate human likeness, while low prediction uncertainty boosts ratings for robot-like appearances. This framework provides a computational basis for algorithmically evaluating and optimizing humanoid robot appearance and behavior.
The uncanny valley effect is translated into four manipulable mathematical variables.
Category ambiguity and appearance-motion mismatch reduce affinity.
High-Voltage Electrostatic Actuators (HVEAs) are emerging as a compelling alternative for haptic interfaces requiring soft, thin, silent, and energy-efficient actuation. This survey reviews four major classes: electrostatic switchable adhesives, dielectric elastomer actuators, soft electrohydraulic actuators, and electrokinetic pumps. It analyzes their mechanisms, bandwidths, force densities, and scalability for rendering cutaneous and kinesthetic feedback, and outlines design constraints and future research directions.
HVEAs offer fast, silent, low-power operation in customizable form factors.
Four classes reviewed: switchable adhesives, DEAs, electrohydraulic actuators, and electrokinetic pumps.
This paper reports a systematic literature review on autonomous UAV route planning for coverage-oriented environmental monitoring. Following the PRISMA 2020 framework, it searches Scopus and Web of Science for studies from 2015 to 2026, focusing on path planning, coverage path planning, and informative path planning. Preliminary analysis of 247 retained studies reveals a concentration on coverage formulations, multi-UAV coordination, and energy-aware optimization, with fewer addressing weather, uncertainty, or obstacles. Most studies rely on simulation validation, indicating a simulation-to-reality gap, and recent work shows growing interest in reinforcement learning, hybrid optimization, and geometry-aware planning.
Systematic review of 562 records (2015-2026), 247 retained for full-text assessment.
Focus on coverage-oriented path planning, multi-UAV coordination, and energy-aware optimization; limited attention to weather, uncertainty, and obstacles.
A robust polarization-aware differentiable path tracing method is proposed, using path replay and local caching to estimate unbiased gradients, overcoming rank deficiency in polarimetric operators for stable inverse rendering.
Existing differentiable rendering ignores polarization, losing geometric and material constraints.
Forward polarization simulation is well-defined, but reverse-mode differentiation suffers from rank-deficient operators.
We present AffectFlow-DINO, a multi-task learning system for the 11th ABAW challenge that extends a standard deterministic architecture with a conditional rectified-flow head to model the inherent ambiguity of in-the-wild facial behavior. The system jointly estimates continuous valence-arousal, classifies eight facial expressions, and detects twelve Action Units. Extensive ablation studies show that rectified-flow decoding consistently improves deterministic prediction, and post-hoc threshold calibration effectively recovers performance on rare classes. The final model achieves P_MTL=1.177, substantially outperforming the official baseline of 0.45.
AffectFlow-DINO uses a conditional rectified-flow generative distribution for uncertainty-aware one-to-many predictions.
The system jointly handles valence-arousal regression, facial expression classification, and Action Unit detection.
This paper introduces JITOMA (Just-In-Time On-demand Memory Activation), a closed-loop framework that unifies task reasoning, perception, and memory to combat perceptual saturation in long-horizon robotics. It uses a task heatmap to filter observations and an LLM to dynamically activate relevant anchors, reducing computational overhead while maintaining stable performance. The authors also present JITOMA-Bench for evaluation.
Conventional 3D scene graph pipelines suffer from perceptual saturation due to exhaustive mapping.
JITOMA uses a task heatmap for observation filtering and LLM for on-demand anchor activation.
A human-in-the-loop framework combining active learning and dual-loss optimization reduces annotation effort for laparoscopic video segmentation by 50%. It uses a foundation model to generate temporally consistent CAMs, with weak supervision on video-level labels and image-level mask loss on human-corrected annotations from active learning. Iterative pseudo-mask refinement eliminates the need for dense initial annotations.
Reduces surgical video annotation effort by 50% using active learning and weak supervision.
Employs a foundation model to generate temporally consistent class activation maps (CAMs).
Static deepfake detectors suffer drastic AUC drops of 45-50% on real-world content due to being trained once against a moving generative frontier. BitMind Forensics (BMF), trained via the Bittensor SN34 open adversarial competition, continuously refreshes its training distribution and achieves strong results across 19 public datasets, including robustness to JPEG compression and downscaling, and improvements over time on unseen generators.
Static detectors fail in the wild with 45-50% AUC drop due to mismatch with evolving generators.
BMF uses continuous adversarial training via Bittensor SN34 to adapt to new deepfake techniques.
A new study systematically compares pretrain-finetuning (PFT) and joint training (JT) paradigms for self-supervised learning, finding JT superior in data efficiency and low-label settings, while PFT is more reliable in specialized domains.
The study compares eight SSL methods across diverse vision tasks with varying labeled data ratios.
Joint training optimizes self-supervised and supervised losses simultaneously, showing robustness in low-label regimes.
MGFace is a mask-gated face identification pipeline that predicts the mask status of a query face and conditionally routes similarity computation: global embedding matching for unmasked queries, and mask-aware patch-level re-ranking only for masked queries. On the extended LFW-Mask dataset, it achieves over 80% accuracy with FaceNet and over 90% with ArcFace, while reducing query time by approximately 20x compared to a prior EMD-based method.
Conditionally routes similarity based on predicted mask status, avoiding unnecessary fine-grained computation
Activates patch-level re-ranking only for masked queries, focusing on upper-face regions
A Transformer-based Masked Autoencoder learns representations for unsupervised steel surface defect recognition. Pretraining masks 75% of image patches, a lightweight decoder reconstructs them, and an auxiliary defect localization objective is jointly trained. Decoder achieves SSIM 0.92, MSE 0.47, and clustering yields 91.3% Hungarian matched accuracy on six defect categories.
Masked autoencoder learns defect representations from unlabeled steel surface images
75% of patches masked during pretraining; decoder reconstructs, encoder jointly trained with localization
Boogu-Image-0.1 is an open-source family of unified multimodal understanding and generation models including Base, Turbo, Edit, and Edit-Turbo variants. It delivers competitive performance in text-to-image generation, fast inference, instruction-based editing, and bilingual text rendering. Through targeted improvements in model understanding, data quality, and training pipelines, coupled with agentic inference-time scaling, it achieves results approaching leading closed-source systems with only 208.62 million unique images and a training cost of approximately $400K.
Boogu-Image-0.1 is an open-source unified multimodal model family with multiple variants
Competitive in text-to-image generation, fast inference, instruction editing, and bilingual rendering
A new method called C-Norm addresses poor AI performance in cervical cancer screening by normalizing cell distribution in TCT images. It decouples abnormal and normal cells and re-synthesizes them for uniform distribution, then uses a hybrid YOLOv12-DINOv3 architecture for detection. Experiments show state-of-the-art results.
C-Norm normalizes cell distribution by decoupling and re-synthesizing abnormal and normal cells.
Integrates YOLOv12 with DINOv3 for improved feature representation.
We propose Samba, a hybrid Mamba architecture for audio-visual navigation. It uses an adaptive selection-enabled Mamba State Encoder (M-SE) to replace conventional GRUs for temporal aggregation, and an Audio Mamba Encoder (AME) to address limitations of convolutional operators in capturing global time-frequency dependencies in spectrograms. Experiments show an 11.3% improvement in success rate on Matterport3D and even better performance on Replica, with lower computational cost. Accepted at IEEE SMC 2026.
Proposes Samba, a hybrid Mamba architecture with M-SE replacing GRU and AME for spectrograms
Improves navigation success rate by 11.3% on Matterport3D, with greater gains on Replica
Real-world traffic data exhibit heterogeneous spatial correlations and nonlinear temporal dynamics, posing challenges for spatio-temporal forecasting. Existing approaches focus on graph, attention, and decomposition architectures but neglect the underlying nonlinear function approximator. STKAN introduces Taylor-polynomial Kolmogorov-Arnold Network modules into spatial and temporal token mixing. It constructs high-level spatial representations via a learnable soft node-group assignment, applies group-wise spatial mixing, models temporal dependencies, and uses self-attention for long-range interactions. Experiments on five traffic benchmarks show competitive performance, outperforming an MLP-based variant, suggesting that nonlinear function approximator design complements architectural innovation.
STKAN integrates Taylor-polynomial KAN modules into spatio-temporal forecasting.
Uses soft node-group assignment and group-wise spatial mixing for spatial features.
Knowledge tracing (KT) predicts student performance by modeling evolving knowledge states. Existing methods treat interactions as a unified process, ignoring phase-specific learning. We propose Phase-Aware Knowledge Tracing (PAKT), which decomposes interactions into ability and proficiency phases. A multi-branch Transformer with type-aware readout captures phase-specific and holistic states. Causal analysis reveals confounding bias in phase-agnostic models. On six benchmarks, PAKT achieves up to 1.33% AUC improvement, averaging 0.82%.
Current KT models overlook distinct learning phases like ability-building vs. proficiency.
PAKT decomposes student interactions into ability and proficiency phases.
Global station weather forecasting (GSWF) is crucial for localized and extreme weather prediction. Existing methods rely heavily on short-term patterns, struggling with chaotic dynamics and partial observations. We propose TSSM, a triaxial state space model with history-enhanced temporal-variable-historical paradigm. It incorporates period-aligned historical data to capture long-term, large-scale periodic patterns. TSSM achieves SOTA on Weather-5K with 10% accuracy gain and 61% extreme event improvement. It excels in long-horizon forecasting (37.5% gain at 240h) and iterative settings (103.5% gain at 48h×5). Under 80% missing observations, TSSM retains >90% performance, demonstrating robustness for real-world deployment.
TSSM uses period-aligned historical data to enhance short-term weather forecasting.
Achieves state-of-the-art on Weather-5K dataset with 10% accuracy and 61% extreme event gains.
A new framework treats the decision of when to invoke a large language model (LLM) in streaming inference as a risk-based sequential stopping problem. The authors prove six theoretical results covering minimum inter-event times, optimality of threshold policies, and regret bounds. Empirical tests on turbofan degradation data show that anomaly-score-driven risk functions outperform baseline methods by an order of magnitude in Pareto AUC.
Formal treatment of LLM invocation timing in streaming systems using risk-based sequential stopping.
Six theoretical results including regret bounds and convergence guarantees.
Parameter decomposition (PD) decomposes neural networks into interpretable components but is computationally expensive for large models. The proposed targeted PD (tPD) introduces a high-rank catch-all component to handle non-target data, enabling efficient recovery of circuits for specific inputs. tPD extracts a CSS-only submodel from a 4-block transformer using 7% of the FLOPs of published decomposition, and surgically ablates memorized sequences in a 12-block transformer with negligible side effects. Accepted at the Mechanistic Interpretability Workshop, ICML 2026.
Targeted PD (tPD) isolates components for specific inputs via a high-rank catch-all component
Extracts CSS-only submodel from 4-block transformer at 7% FLOPs of existing decomposition
This paper presents a framework to analyze information discarded by machine learning models when inputs exhibit Lie group symmetries. It defines null fibers and stabilizers to measure the symmetry invisible to a model, and uses the Peter-Weyl theorem for a spectral characterization on compact groups. Efficient computation via Newton iteration is demonstrated. Applications to data masking, model fingerprinting, and privacy-preserving computation are experimentally validated on molecular property prediction under SO(3) and spherical image classification under the Möbius group PSL(2, C). The framework applies uniformly to classical neural networks and variational quantum circuits.
Introduces null fibers and stabilizers to quantify a model's insensitivity to Lie group symmetries.
Provides spectral characterization via Peter-Weyl theorem for compact groups and efficient computation via Newton iteration.
Federated Learning (FL) enables privacy-preserving collaborative training across distributed data, but lacks model transparency. Explainable AI (XAI) addresses opacity. Their intersection, Federated Explainable AI (FedXAI), is systematically reviewed in this survey, highlighting explainability's shift from post-hoc tool to integral FL component. A taxonomy classifies methods by role, model type, scope, integration level, FL settings, and data heterogeneity. Evaluation practices lack standardized benchmarks. Key challenges include non-IID data, security threats, communication efficiency, continual learning, and domain knowledge integration.
FedXAI integrates explainability throughout the FL lifecycle, supporting aggregation, personalization, robustness, and coordination.
The survey proposes a taxonomy organizing FedXAI methods by explainability role, model type, explanation scope, integration level, FL settings, and data heterogeneity.
A new lightweight training strategy for deep learning models decouples feature extraction from classifier optimization, drastically reducing training time and energy consumption with minimal accuracy loss, as tested on multiple architectures and medical datasets.
Novel decoupled training strategy adapts normalization layers and precomputes features once.
Achieves significant reduction in training time and CO2 emissions.
This paper traces, with explicit numerical values, how PyTorch's automatic differentiation engine computes gradients for Physics-Informed Neural Network training, including physics derivative computation and parameter gradients. Using a 1-3-3-1 multilayer perceptron and a simple ODE, it details the computational graph, reverse-mode backward traversal, and the graph-on-graph mechanism for correct differentiation through physics-informed residuals.
Demonstrates two-level differentiation in PINNs: physics derivative ŷ' and parameter gradient ∇θL
Traces complete pipeline with a 1-3-3-1 MLP and ODE initial value problem
A new system called Mycelium enables networked intelligence by connecting researchers and AI agents in a shared workspace, automatically routing observations and hypotheses to relevant team members. Tested in a biological multi-omics campaign, it turned local findings into cross-expert constraints and experimental designs.
Most AI-for-science systems focus on scaling individual reasoning, but complex problems require team collaboration.
Mycelium creates an active shared workspace that captures and routes context among humans and AI agents.
cayleyR is an R package for solving permutation puzzles by detecting cycle intersections in Cayley graphs. It uses an iterative bidirectional search and distance-guided bridge selection. The package targets the TopSpin(n,k) puzzle and leverages C++ indexing with optional Vulkan GPU acceleration. It is available on CRAN.
cayleyR solves permutation puzzles via cycle intersection detection in Cayley graphs
Uses iterative bidirectional search with distance-guided bridge selection
A technical report from arXiv introduces Oracle Agent Memory, a database-native memory system built on Oracle Database for long-horizon AI agents. It achieves 93.8% accuracy on LongMemEval while using 10.7x fewer tokens compared to flat-history baselines. The system addresses memory lifecycle, layered architecture with scope control, and evaluation methodology combining task accuracy with memory-specific metrics.
Agent memory is critical for long-horizon AI agents to retain state, user preferences, and procedural knowledge.
Oracle Agent Memory is built on Oracle Database with a lifecycle covering ingestion, extraction, consolidation, retrieval, summarization, and revision/removal.
Small language models struggle with molecular property prediction due to structural blindness. A new framework called Context-Augmented Prompting integrates GNN tools to provide predictive hints and explanatory subgraphs, achieving up to 74% relative improvement on Tox21, though a gap with specialized GNNs remains.
SLMs miss graph-topological cues in SMILES sequences.
Proposed framework uses GNN expert for hints and subgraph extraction.
This survey reviews self-improving autonomous agents transitioning from prototypes to deployed systems. It introduces a system-level framework modeling an agent as a foundation model coupled with an operational scaffold (prompts, memory, tools, control logic). Self-improvement is formalized as a self-induced update operator that updates model parameters or scaffold components. The paper categorizes prior work by update target and driving signals, and discusses applications, evaluation, and open challenges.
Self-improving agents are moving from research to deployment with minimal human input
The framework models agents as foundation models combined with an operational scaffold
This paper presents a probabilistic extension of neuro-symbolic AGI based on Belnap's Typed Intensional FOL (IFOL_B). By integrating Nilsson's probability structure, it computes probabilities for unknown sentences, introducing global and local symmetry transformations to preserve knowledge integrity and enable real-time decision-making. Neural networks compute the probability density function via Shannon's maximum information entropy.
Combines neural learning and symbolic reasoning to overcome limitations of purely neural systems.
Introduces Nilsson's probability structure for computing probabilities of unknown sentences.
Large language models produce chain-of-thought reasoning that appears logically sound but may not genuinely depend on its stated premises. This paper introduces interventional grounding audits, a black-box step-level test of premise dependency that substitutes a predicate in a single premise and checks for changes in the normalized conclusion of each reasoning step. Evaluated on ProntoQA with GPT-4o, the method achieves F1=0.806 for detecting proof-tree dependencies, significantly outperforming a self-consistency baseline (F1=0.343). Notably, 66% of correctly solved problems contain at least one step insensitive to a direct proof-tree dependency, revealing a 'right answer, wrong reasoning' signal.
Interventional grounding audits test premise dependency at the step level by substituting predicates in a black-box manner.
On ProntoQA, the method achieves F1=0.806 for proof-tree dependencies, outperforming self-consistency (F1=0.343).
Researchers propose SPINE, an agentic framework that automates debugging and deployment of bimanual robots, reducing reliance on expert calibration. In tests, SPINE improved success rates and reduced time-to-teleoperation compared to manual methods.
SPINE uses multi-agent workflows for robot profile building and iterative debugging.
Novices using SPINE outperformed experts on DOBOT X-Trainer, achieving 100% success.
OriginBlame is a record- and token-level data provenance system that precisely resolves data removal requests to individual training records, reducing over-deletion from 101x to 1.3x on Wikipedia data. Integration adds 1.3-4.0% throughput overhead (HuggingFace) and 2.1-19.0% (Datatrove). On a 1.7B model, provenance-based forget sets improve unlearning by 42% over random baselines.
OriginBlame provides record- and token-level data provenance for AI training datasets.
It reduces over-deletion from 101x to 1.3x on Wikipedia data.
Service organizations supporting critical infrastructure face a structural mismatch: tightening uptime requirements while maintenance models and technician capacity lag. AI offers anomaly detection for condition-based maintenance, prescriptive guidance for consistent technician performance, and requires operational transformation to succeed.
Anomaly detection enables proactive maintenance by alerting technicians to behavioral drifts before failures occur.
Mira Murati's Thinking Machines Lab Inc. today launched its first foundation model with the release of Inkling, making its full open weights available to developers so they can fine-tune it as they wish. Inkling is a mixture-of-experts model with 975 billion parameters (41B active) trained on 45 trillion tokens of text, image, audio and video, capable of reasoning across all four modalities but outputting only text. It features "thinking effort" controls and uncertainty flagging to reduce hallucinations. The model is fine-tunable via the Tinker API and aims to provide a Western open-source alternative to Chinese AI models. Thinking Machines plans to generate revenue through the Tinker platform rather than per-token API access, potentially disrupting current AI business models.
Thinking Machines releases Inkling, a 975B-parameter open-weights model (41B active).
Trained on 45T tokens across modalities; outputs text only.
A VentureBeat Pulse Research survey of 101 enterprises reveals that agent orchestration is consolidating on model-provider platforms, with Anthropic Claude leading at 40%. However, 71% admit that a quarter or fewer of their deployed 'agents' are true multi-step workflows, and only 10% have crossed the halfway mark. Enterprises plan hybrid control planes to avoid vendor lock-in, but real-time cost control remains immature.
Anthropic Claude is the primary orchestration platform for 40% of enterprises, more than double any rival.
71% of enterprises say a quarter or fewer of their deployed 'agents' are truly orchestrated multi-step workflows.
AWS Marketplace has introduced AI-assisted product listings to help partners create comprehensive listings using existing assets, optimize for SEO, and adapt to the growing use of enterprise agents. The agentic AI category has grown from 900 to over 3,400 partners in under a year.
AWS Marketplace launches AI-assisted product listings to reduce manual data entry and improve SEO.
Agentic AI category becomes fastest-growing, with partners surging from 900 to over 3,400.
Opposition to AI data centers is a growing political issue in the US, but it may distract from the larger threat: the concentration of wealth and power in AI companies. This article argues that while data centers have local costs, AI's real impact is the takeover of entire industries and political influence. Solutions include regulation, taxation, and a public AI ecosystem.
Opposition to data centers diverts attention from AI companies' power concentration.
AI firms aim to control entire industries like education and healthcare.