All Updates
Page 1848 of 1900
February 18, 2026
WD Raises $3.09B Selling SanDisk Shares
Western Digital plans to raise $3.09 billion by selling shares in its flash memory subsidiary SanDisk, spun off nearly a year ago. Shares priced at $535-$555, up to 9.4% discount from $590.59 close.
X-MAP Profiles Misclassifications in Spam Detection
X-MAP is an explainable framework combining SHAP attributions and NMF to create topic profiles for correctly classified spam/phishing vs. legitimate messages. It detects misclassifications via Jensen-Shannon divergence from these profiles. Experiments achieve 0.98 AUROC and recover 97% of false rejections when used as a repair layer.
WGAN Boosts Synthetic Population Diversity
New method uses joint WGAN with gradient penalty to synthesize populations from multi-source data, tackling diversity and feasibility issues. Introduces regularization term for generator loss, outperforming baselines in recall (+7%), precision (+15%), and overall similarity (88.1 vs 84.6). Enhances agent-based models in transportation and urban planning.
WAC Boosts Web Agents with World Models
WAC integrates multi-agent collaboration where an action model consults a world model for strategic guidance on web tasks, grounding suggestions into executable actions. It employs a two-stage deduction chain with consequence simulation and judge model scrutiny for risk-aware action correction. Experiments show 1.8% gains on VisualWebArena and 1.3% on Online-Mind2Web.
Trace Rewriting Blocks LLM Distillation Theft
Researchers introduce methods to rewrite teacher model reasoning traces, deterring unauthorized knowledge distillation while embedding verifiable watermarks. Techniques include LLM-powered rewriting and gradient-based approaches that preserve answer correctness. Experiments demonstrate strong anti-distillation effects with maintained or improved teacher performance and reliable watermark detection.
Tech Billionaires Flock to Delhi AI Summit
Silicon Valley leaders from Google, Anthropic, and OpenAI will attend India's week-long AI Impact Summit hosted by PM Narendra Modi. The event gathers thousands of tech executives, officials, and AI safety experts alongside global south leaders like those from Kenya and Indonesia. Discussions focus on wresting control over rapidly advancing AI technology.
Secure Energy-Efficient Agentic AI Wireless Networks
This arXiv paper introduces a secure wireless agentic AI network with a supervisor AI agent assigning others for cooperative reasoning while unselected agents jam eavesdroppers. It formulates an energy minimization problem optimizing agent selection, BS beamforming, and transmission power under latency and accuracy constraints. Proposed ASC and LAW schemes reduce energy by up to 59.1%, validated on a Qwen-based system with strong benchmark performance.
Ruva: Transparent On-Device Personal AI Graph Reasoning
Ruva introduces a 'Glass Box' architecture for Personal AI using Personal Knowledge Graphs, allowing users to inspect and precisely redact AI knowledge. It replaces vector databases with graph reasoning to eliminate hallucinations, ensure accountability, and enable the 'Right to be Forgotten.' A demo and project are available online.
ResearchGym: AI Agents Research Benchmark
ResearchGym introduces a benchmark with five containerized environments from ICML, ICLR, and ACL papers, totaling 39 sub-tasks, where agents propose hypotheses and run experiments to beat baselines. GPT-5-powered agents show a capability-reliability gap, succeeding in only 6.7% of evaluations and completing 26.5% of sub-tasks. It identifies key failure modes like impatience and poor resource management, while occasionally achieving SOTA results unreliably.
Reasoning LLMs Beat Conversational in Risky Choices
A study of 20 LLMs reveals clustering into rational reasoning models (RMs), insensitive to framing and order, and less rational conversational models (CMs) with human-like biases and a description-history gap. RMs match rational agents across explicit and experience-based prospects. Mathematical reasoning training differentiates RMs from CMs.
R2U-Net Hits 0.900 DSC in Brain Tumor Segmentation
Introduces Attention-Gated Recurrent Residual U-Net (R2U-Net) Triplanar model for glioma segmentation, achieving 0.900 Dice Score on BraTS2021 Whole Tumor. Integrates residual, recurrent, and attention mechanisms for efficiency. Extracts features for survival prediction with 45.71% accuracy.
Panini: Continual Learning via GSW Memory
Panini proposes a non-parametric continual learning framework for LLMs using Generative Semantic Workspaces (GSW), an entity- and event-aware QA network that consolidates experiences without updating the base model. It outperforms RAG baselines by 5-7% on six QA benchmarks while using 2-30x fewer tokens and reducing unsupported answers. Open-source code is available on GitHub.
Novel da Costian-Tarskian Ontology Heterogeneity Approach
This arXiv paper proposes da Costian-Tarskianism, blending Carnapian-Goguenism with da Costa's tolerance principle and Tarski's consequence operators for ontological heterogeneity. It introduces extended consequence systems augmented with ontological axioms and extended development graphs for relating ontologies via morphisms, fibring, and splitting. The work discusses implications for applied ontology.
New Model Quantifies LLM Benchmark Validity
Presents structured capabilities model to extract interpretable LLM capabilities from benchmarks, addressing construct validity. Outperforms latent factor models on fit and scaling laws on prediction using OpenLLM Leaderboard data. Combines scaling laws and latent factors by separating model scale from capabilities.
Memory & Planning Excel in Dynamic Navigation
This arXiv paper explores memory strategies for spatial navigation in non-stationary environments with uncertain sensing in a foraging task. It compares simple to sophisticated agents, finding hybrid architectures with episodic memories and on-the-fly planning most efficient for exploration, search, and path optimization. Advanced agents substantially outperform minimal-memory ones as task difficulty increases, provided uncertainty is manageable.
Hybrid Abstention Boosts LLM Reliability
This arXiv paper introduces an adaptive abstention system for LLMs that dynamically adjusts safety thresholds using contextual signals like domain and user history. It features a multi-dimensional detection architecture with five parallel detectors in a hierarchical cascade, reducing latency and false positives. Evaluations show strong performance in sensitive domains like medical advice.
EduEVAL-DB Dataset for AI Tutor Evaluation
EduEVAL-DB introduces a dataset of 854 explanations for 139 ScienceQA questions across K-12 subjects, with one human-teacher and six LLM-simulated teacher explanations. It features a pedagogical risk rubric covering factual correctness, depth, focus, appropriateness, and bias, annotated via semi-automatic expert review. Preliminary benchmarks compare Gemini 2.5 Pro against fine-tuned Llama 3.1 8B for risk detection on consumer hardware.
EAA Automates Microscopy with VLM Agents
Experiment Automation Agents (EAA) is a vision-language-model-driven system that automates complex microscopy workflows in materials characterization. It combines multimodal reasoning, tool actions, and long-term memory for autonomous or user-guided experiments. Demonstrated at Advanced Photon Source, it handles focusing, feature search, and data acquisition to boost efficiency.
Common Belief Defies KD4: New Axioms
Contrary to common belief, common belief is not KD4 under KD45 individual beliefs, retaining only D and 4 properties plus shift-reflexivity C(Cφ → φ). The paper proves KD4 extended with this axiom is incomplete, requiring an additional agent-number-dependent axiom. This fully characterizes common belief, settling a long-open problem.
AI Predicts Invoice Dilution with Leakage-Free XGBoost & KAN
This ArXiv paper proposes an AI/ML framework to predict invoice dilution in supply chain finance, mitigating non-credit risks and margin losses. It employs leakage-free two-stage XGBoost, Kolmogorov-Arnold Networks (KAN), and ensemble models trained on production data across nine transaction fields. The method supports real-time dynamic credit limits, reducing reliance on buyer's irrevocable payment undertakings (IPU).