All Updates

Page 570 of 1667

May 19, 2026

📄
ArXiv AI73d ago

Skim: Speculative Execution Framework for Faster Web Agents

Skim is a new speculative execution framework that optimizes web agents by bypassing heavy model inference for predictable website tasks. It uses offline profiling to match queries to templates, significantly reducing latency and costs without sacrificing accuracy.

#web-agents#llm-optimization#automation
📄
ArXiv AI73d ago

Scalable Uncertainty Reasoning in Knowledge Graphs

This research proposes a modular framework to handle uncertainty in knowledge graphs, addressing imprecise attributes, probabilistic triples, and incomplete schemas. It utilizes algebraic, logical, and geometric approaches to maintain semantic precision while ensuring computational tractability.

#knowledge-graphs#semantic-web
📄
ArXiv AI73d ago

PRISMat: Efficient Permutation-Invariant Material Generation Model

PRISMat is a new, cost-effective model designed for material discovery that outperforms LLMs in crystal slab generation. By addressing the inefficiencies of over-parameterized language models, it significantly reduces error rates in predicting critical surface properties.

#material-science#generative-ai
📄
ArXiv AI73d ago

MetaKGEnrich: Automating Knowledge Graph Repair for LLMs

MetaKGEnrich is an automated pipeline that enables LLMs to perform self-directed knowledge repair by identifying sparse graph regions and retrieving external evidence. The system significantly improves answer accuracy across multiple benchmarks by combining topological diagnosis with GraphRAG.

#knowledge-graph#metacognition#rag
📄
ArXiv AI73d ago

LLMs Struggle to Translate Counterparty Modeling into Strategic Bargaining

Research indicates that while LLMs can accurately model a counterparty's preferences, they fail to leverage this information for strategic advantage in bargaining. Agreements are often driven by surface-level anchors rather than actual utility optimization.

#llm-agents#negotiation-strategy#utility-optimization
📄
ArXiv AI73d ago

LinAlg-Bench Reveals Structural Failure Modes in LLM Math

LinAlg-Bench is a new diagnostic benchmark that evaluates 10 frontier LLMs on linear algebra tasks. It identifies a critical behavioral threshold at 4x4 matrix scales where models shift from arithmetic errors to structured hallucinations and computational abandonment.

#llm-reasoning#benchmarking#linear-algebra
📄
ArXiv AI73d ago

CBEA: Solving Commitment Failures in Personalized Language Systems

This research introduces Contract-Bounded Evidence Activation (CBEA) and Lexicographic Commitment Validation (LCV) to prevent language models from making unreliable commitments. By bounding evidence and validating structured outputs, the system significantly reduces failures in long-context and memory-intensive applications.

#llm-reliability#rag#memory-systems
📄
ArXiv AI73d ago

ANNEAL: Governed Symbolic Repair for Persistent LLM Agent Faults

ANNEAL is a neuro-symbolic agent that fixes recurring execution errors by performing governed symbolic edits on process knowledge graphs. Unlike standard fine-tuning or prompt engineering, it ensures structural reliability through validation, guardrails, and deterministic rollbacks.

#neuro-symbolic#agent-reliability#knowledge-graphs
📄
ArXiv AI73d ago

AI Agent Automates Laboratory Protocols via Natural Language

Researchers have developed an AI agent architecture integrated into the Experiment Orchestration System (EOS) that allows scientists to create and manage lab protocols using natural language. The system features an agentic loop with automated validation and error correction, achieving a 97% success rate in protocol generation.

#agentic-workflow#scientific-discovery
🐯
虎嗅73d ago

Science is aging: Innovation narrowing in academia

A study in Science reveals that as the scientific workforce ages, research is becoming more conservative and reliant on older literature. This 'imprinting effect' and the role of senior scientists as gatekeepers are hindering disruptive innovation.

#academic-innovation#data-analysis#research-methodology
⚛️
量子位73d ago

Qwen 3.7 Max Preview Released: Top-Tier Performance

Alibaba has launched the Qwen 3.7 Max preview, marking a significant milestone in their model development. The release demonstrates industry-leading capabilities in both text and vision domains.

#chinese-llm#multimodal#model-release
🇭🇰
SCMP Technology73d ago

Xpeng Enters Robotaxi Market to Challenge Tesla FSD

Chinese EV maker Xpeng has officially begun mass production of L4 autonomous robotaxis. This move marks a direct escalation in the competition against Tesla's FSD software in the physical AI and self-driving space.

#autonomous-driving#robotaxi#physical-ai
📲
Digital Trends73d ago

Apple Watch may add advanced blood pressure sensing

Apple is reportedly developing a more sophisticated blood pressure monitoring feature for its upcoming high-end Apple Watch models. This builds upon existing hypertension warning capabilities to provide more granular health data.

#wearables#biometrics#digital-health
🔥
36氪73d ago

Tsinghua and Inbo Tech partner on GPU cloud

Tsinghua University Shenzhen International Graduate School partnered with Inbo Tech to launch a GPU cloud platform. The collaboration aims to support academic research with dedicated AI computing infrastructure.

#gpu#cloud-computing#academic-partnership
⚛️
量子位73d ago

Baidu Robotaxi hits 350k weekly orders, achieves city-level profit

Baidu's autonomous driving service, Apollo Go, has reached a new milestone with over 350,000 weekly orders. CEO Robin Li confirmed that the service has achieved profitability in at least one city, marking a significant step for commercial robotaxi operations.

#robotaxi#autonomous-driving#commercialization
🐯
虎嗅73d ago

Meta and Tesla Use Human Data to Train AI Agents

Tech giants like Meta and Tesla are increasingly collecting granular human behavioral data—from mouse movements to physical labor actions—to train AI agents to perform human-like tasks. This trend highlights a shift from human-AI collaboration to the 'distillation' of human skills for automation.

#embodied-ai#data-collection#robotics
🔥
36氪73d ago

SportVision raises funding for AI sports coaching

SportVision, an AI sports tech company, raised Angel+ funding led by GL Ventures. They are building an 'AI Sports Copilot' that provides real-time movement analysis and coaching for racket sports.

#computer-vision#sports-tech#ai-copilot
🔥
36氪73d ago

Apple Announces WWDC 2025 Event Schedule

Apple has officially scheduled its annual Worldwide Developers Conference (WWDC) from June 9 to June 13. The event will focus on platform updates, including new software, developer tools, and significant advancements in artificial intelligence.

#apple-ecosystem#developer-conference#ai-integration
🐯
虎嗅73d ago

Musk loses first legal battle against OpenAI

A federal jury dismissed Elon Musk's lawsuit against OpenAI, citing the statute of limitations regarding the company's transition from a non-profit to a commercial entity. The court did not rule on the core ethical allegations, leaving the debate over AI's mission versus commercialization unresolved.

#legal-battle#ipo#corporate-governance
🐯
虎嗅73d ago

SMIC acquires remaining stake in SMIC North

中芯国际把中芯北方收回来了,但考验才开始 本文来自微信公众号: 心智观察所 ,作者:心智观察所 5月11日傍晚,上交所公告,中芯国际发行股份购买资产的事项审议通过。 当天科创50指数大涨近5%,Wind半导体指数涨幅接近6%,沪指顺势站上4200点。中芯国际A股微涨1.84%,报122.52元,港股一路拉升4.43%。 被这场行情点燃的,是中国半导体产业过去十三年里最大的一笔并购:中芯国际以约406.01亿元的对价,向五家股东发行股份,把子公司中芯北方

#semiconductor#supply-chain#merger
Page 570 of 1667