All Updates
Page 725 of 1667
May 6, 2026
AI Translates Pig Farming Expertise for Efficiency
iFlytek and Heguan Technology use large models to 'translate' traditional breeding experience into smart pig farming systems. This boosts quality and efficiency in pig raising. The initiative injects new momentum into upgrading traditional industries.
Anthropic Launches Finance AI Agents
Anthropic has unveiled new AI agents designed for a broader range of financial services tasks. This initiative is part of the company's strategy to attract Wall Street clients. The news was reported by Bloomberg's Avril Hong.
AMD 16-Core PRO X3D CPU Spotted in Database
AMD's unreleased Ryzen 9 PRO 9965X3D has appeared in the PassMark database. It marks the first 3D V-Cache processor in the PRO series and the line's first 16-core/32-thread model. The CPU was previously spotted in logistics records.
Agentic AI Shifts CPU/GPU to 1:1 Ratio
AMD CEO Lisa Su states agentic AI is reshaping server CPU demand, with hyperscalers expanding EPYC deployments for orchestration and data tasks. CPU needs complement GPUs, evolving data center ratios from 1:4/8 to 1:1 or higher. Server CPU market TAM could see 35% CAGR to 2030.
DeepSeek V4 Optimized for Huawei Ascend
DeepSeek launched next-generation V4 AI models optimized for Huawei's Ascend chips and software tools. This exemplifies the growing divergence in China-US AI ecosystems amid the tech war. China's self-reliance policy is being reinforced by the Communist Party.
Meta Develops Personalized AI Agents for Billions
Meta is creating highly personalized AI assistants for its billions of users to handle daily tasks. The effort comes under investor scrutiny due to escalating AI spending costs.
China Pharma's AI-Driven Decade Growth
China's pharma industry scaled up with AI, biotech fusion amid policies like '14th Five-Year Plan'. Revenue stabilized post-fluctuations, exports doubled, shifting to quality-global innovation. Review times slashed, ICH joined for standards.
Proven Transparent Governance for AI Workflows
Researchers present a machine-checked formalization of effect-transparent governance for AI workflow architectures using Interaction Trees in Rocq. They prove governance preserves Turing completeness, oracle expressivity, and semantic transparency without reducing computational power. Seven key properties establish governance as orthogonal to expressivity.
PERSA: RLHF for Professor-Style LLM Feedback
PERSA introduces an RLHF pipeline to align LLMs with a professor's personalized feedback style in programming education without compromising accuracy. It combines supervised fine-tuning, reward modeling from pairwise preferences, and PPO, using parameter-efficient updates on top transformer blocks. Evaluations on APPS, PyFiXV, and CodeReviewQA show 96.2% style alignment and up to 100% correctness on Llama-3 and Gemma-2.
Multi-Agent Excels in Hydrodynamics Reasoning
Researchers present a multi-agent system (MAS) prototype using Layer Execution Graph (LEG) to coordinate specialized agents for hydrodynamics tasks, addressing context saturation in single-agent LLM workflows. Specialist agents use strict tool allowlists, consolidators fuse outputs, and reporters synthesize responses with full provenance logging. Benchmarks on 37 queries show 93.6% factual precision, 100% pass rate, and robust performance under parallel runs or data loss.
LLMs Crack New Zarankiewicz Bounds
Researchers used OpenEvolve, an LLM-based evolutionary algorithm, to determine exact values for three Zarankiewicz numbers: Z(11,21,3,3)=116, Z(11,22,3,3)=121, and Z(12,22,3,3)=132. They also established lower bounds for 41 more cases, some nearly matching upper bounds. The method costs under $30 per combination, showcasing affordable LLM-guided math discovery.
LLM System for Explainable LPBF Defect Analysis
This arXiv paper introduces a knowledge-driven LLM-based system for defect diagnosis and mitigation in Laser Powder Bed Fusion (LPBF) manufacturing. It integrates an ontology with 27 defect types, supports fuzzy natural language queries, and features multimodal image assessment. Evaluations show a macro F1 score of 0.808 and substantial inter-rater agreement.
Feature Superposition Explains Emergent Misalignment
Researchers propose a geometric model linking emergent misalignment in LLMs to feature superposition, where fine-tuning amplifies nearby harmful features. Empirical tests across Gemma-2, LLaMA-3.1, and GPT-OSS using sparse autoencoders confirm harmful features cluster closer to misalignment-inducing data. A geometry-aware filtering method cuts misalignment by 34.5%, outperforming baselines.
Clinician-in-Loop AI Stuttering Therapy Agent
Virtual Speech Therapist (VST) automates stuttering assessment using deep learning classification and generates personalized therapy plans via multi-agent LLM reasoning. A critic agent ensures safety and evidence alignment, with clinician feedback for finalization. Expert therapists validated its high-quality outputs; live demo available online.
ClinicBot: Prioritized RAG for Clinical Chatbots
ClinicBot is an AI clinical chatbot that grounds responses in official guidelines via structured extraction and prioritized evidence RAG. It ranks content by clinical significance rather than similarity, reducing hallucinations in medical contexts. A web demo showcases diabetes queries using ADA 2025 standards.
Algebraic Semantics for Governed Execution
This paper introduces an algebraic semantics for governed execution, axiomatized with three governance axioms forming a symmetric monoidal category. It is mechanized in 32 Rocq modules with 454 theorems and extracted to OCaml running as a NIF in BEAM, verified via property-based testing. Key results include capability-indexed composition and a coterminous boundary theorem ensuring expressibility matches governance.
AI Interface Accelerates Battery Research
This study optimizes sodium-ion battery formation protocols using an AI interface between FINALES and Kadi4Mat frameworks. It employs multi-objective batched Bayesian optimization to minimize formation time while maximizing end-of-life performance, reducing experiments needed. The interoperable setup bridges automated systems and human workflows across research centers.
AI Agents for SME ESG Assessment
This arXiv paper introduces an AI-driven framework for ESG performance assessment in European SMEs, using expert-validated baselines from Flash Eurobarometer FL549 survey data. A scalable AI agent system on the n8n platform employs LLMs for automated classification and contextual recommendations. Results show high consistency with human outputs, enabling effective monitoring aligned with the European Green Deal.
2026 AI/ML Roadmap for Smart Manufacturing
This arXiv paper presents a 2026 roadmap for AI and ML in smart manufacturing, tackling challenges like big data complexity and system integration. It is divided into foundations and trends, key applications such as digital twins and robotics, and emerging approaches including physics-informed AI, generative AI, and LLMs. The roadmap guides researchers and practitioners to overcome barriers and drive innovation.
Kid AI Marketing Hype Explodes
Parents and institutions hype children's simple AI-prompted projects as genius-level feats on social media, from Arduino smart cars to Gemini-generated Minecraft clones. The article critiques packaging play as expertise, urging traditional programming over 'AI speed camps.' This reflects meritocracy finding fertile ground in short-video AI kid content.