All Updates
Page 35 of 1339
June 16, 2026
Security leaders urge lifting export controls on Anthropic models
Security experts are advocating for the removal of export controls on Anthropic's Mythos-class models. They argue that current limitations hinder defensive capabilities and put security researchers at a disadvantage.
SERAF: Enhancing Time Series Forecasting with Multimodal Retrieval
SERAF is a new framework that improves time series forecasting by combining numerical data with self-generated textual descriptions. By performing dual retrieval, it overcomes the limitations of traditional similarity-based methods in non-stationary environments.
PrologMCP: Standardized Prolog Tool Interface for LLM Agents
PrologMCP is an open-source server that enables LLM agents to use Prolog for symbolic reasoning via the Model Context Protocol (MCP). It allows agents to delegate complex deductive tasks to a formal solver, significantly improving accuracy on logic-heavy benchmarks compared to standard LLM reasoning.
OSGuard: A New Safety Benchmark for Computer-Use Agents
OSGuard is a dual-granularity benchmark designed to evaluate the safety of computer-use agents by testing both local guardrail decisions and end-to-end execution. It helps identify gaps where agents might achieve task goals through unsafe shortcuts in desktop or web environments.
New Relational Structural Causal Models for Combinatorial AI
This research introduces Relational Structural Causal Models to enable AI systems to reason about interventions and counterfactuals in environments with varying objects. It provides a formal framework for identifying causal queries in unseen scenarios, outperforming non-relational baselines in simulated traffic environments.
New Framework Improves Multimodal Clinical Time-to-Event Predictions
Researchers introduced a foundation model-driven framework to align CT imaging and EHR data for better time-to-event predictions. The study systematically evaluates four fusion strategies to address modality imbalance in clinical settings.
Measuring Trust Dynamics Between AI Agents
This research proposes a behavioral framework to measure trust between AI agents using costly verification in cooperative tasks. It reveals that frontier models like Claude Opus and GPT-5.1 exhibit distinct trust formation and recovery patterns, impacting overall system efficiency.
Geometric Framework Identifies Memory Traces in Neural Networks
Researchers have developed a geometric framework to isolate 'AI engrams,' allowing for the surgical manipulation of specific memories within deep neural networks. This approach enables the composition or erasure of learned knowledge through linear arithmetic without requiring iterative retraining.
Dr-DCI: Scaling Agentic Search via Dynamic Workspace Expansion
Dr-DCI is a new framework that enhances agentic search by dynamically pulling relevant documents into a local workspace. This approach combines the scalability of retriever-based systems with the precision of direct corpus interaction, outperforming traditional search methods.
Defining Good Explanations for LLM Outputs
This research paper proposes a new definition for AI explainability based on counterfactuals and the interlocutor's prior beliefs. It highlights the inherent difficulties in generating meaningful explanations for complex LLM outputs.
Cognitive Debt: The Hidden Risk of Over-Reliance on AI
This research introduces the concept of 'cognitive debt,' describing the accumulation of unverified reasoning obligations when AI replaces rather than complements human cognition. It warns that short-term productivity gains can mask systemic fragility, leading to potential cognitive 'Minsky moments.'
Huawei Celia: Strategic AI integration and model adaptation
Huawei's Celia assistant is undergoing strategic internal optimizations to improve its AI capabilities. The focus is currently on expanding model adaptation across a wider range of hardware devices.
Anthropic forced to suspend Claude Mythos 5 over export controls
Anthropic received a US export control directive requiring the suspension of access to its Mythos 5 and Fable 5 models for all foreign nationals. The company had to disable these products globally while attempting to negotiate with the Trump administration.
Microsoft Debuts Intelligent Terminal with Native AI Agent Integration
Microsoft has released an experimental 'Intelligent Terminal' for Windows 11, featuring a sidebar-integrated AI agent. It supports real-time error detection, command generation, and multi-step task assistance directly within the shell.
Xiaohongshu prepares for confidential Hong Kong IPO filing
Xiaohongshu is reportedly preparing to file for a confidential IPO in Hong Kong by the end of this month. The Shanghai-based social platform is working with financial advisers on what could become one of the city's largest listings.
Google Gemini integrates deeply with mobile OS
Google is shifting its strategic focus away from traditional Android interfaces toward a Gemini-centric mobile experience. This suggests a fundamental change in how users interact with mobile operating systems via AI.
China's RISC-V ecosystem sees rapid diversification
The RISC-V architecture is achieving large-scale commercialization in China. The ecosystem is expanding rapidly across various hardware applications.
SpaceX Starlink: The core of AI strategy funding
Starlink has become the primary driver of SpaceX's financial growth and profitability. Its revenue is now providing essential capital for the company's broader AI initiatives.
DJI launches dual-camera Osmo Pocket 4P cinema camera
DJI has unveiled the Osmo Pocket 4P, the first in its series to feature a dual-camera system. It includes a 1-inch LOFIC sensor and a dedicated 60mm telephoto lens for advanced mobile cinematography.
Alibaba's Internal AI Agent Battle: QoderWork vs Wukong
Alibaba is internally testing QoderWork against the DingTalk-based Wukong agent. The company aims to determine its flagship enterprise AI product to compete with Tencent and ByteDance.