All Updates
Page 605 of 1669
May 15, 2026
Preping: Building Agent Memory via Proposer-Guided Synthetic Practice
Preping is a new framework that enables AI agents to build procedural memory before encountering real-world tasks by using self-generated synthetic practice. By employing a proposer-solver-validator loop, it ensures high-quality memory construction while significantly reducing deployment costs compared to traditional online methods.
PolitNuggets: Benchmarking Agentic Discovery of Long-Tail Political Facts
PolitNuggets is a new multilingual benchmark designed to evaluate how agentic Large Reasoning Models (LRMs) discover and synthesize long-tail political facts. It features a dataset of 400 global elites and introduces FactNet to measure discovery efficiency and factual accuracy.
New GraphRAG Framework Aligns LLM Agents with Social Values
Researchers proposed a value-based framework using GraphRAG to convert abstract principles into actionable instructions for LLM agents. This approach improves decision-making in complex dilemmas by incorporating psychological theories like Maslow's Hierarchy of Needs.
New 2D Framework for AI Agent Architecture Design
This research introduces a two-dimensional classification framework for AI agents, mapping cognitive functions against execution topologies. It identifies 27 distinct design patterns to help practitioners standardize agent architecture and selection.
NERVE: Network-Aware Bilinear Tokenization for Brain Connectivity
NERVE is a self-supervised learning framework that improves brain functional connectivity (FC) representation by partitioning matrices into network-aware patches. It uses a structured bilinear factorization to preserve network identity while reducing computational complexity from quadratic to linear.
MIGP Optimizes Personalized Nutrition with Integer Serving Constraints
Researchers introduced Mixed Integer Goal Programming (MIGP) to solve the limitations of traditional diet optimization, such as impractical fractional servings and infeasible hard constraints. The method uses integer variables for realistic serving sizes and goal programming to balance nutrient targets effectively.
Invisible Orchestrators Create Safety Risks in Multi-Agent Systems
This research reveals that hidden orchestrators in multi-agent AI systems can cause collective dissociation and internal-state distortion. While output performance remains high, the study warns that behavior-based evaluations are insufficient to detect these critical safety risks.
GraphBit: Deterministic DAG-based Orchestration for Reliable AI Agents
GraphBit is a new agentic framework that replaces prompted orchestration with a Rust-based engine using directed acyclic graphs (DAGs). It eliminates common issues like hallucinated routing and infinite loops, achieving superior performance and reproducibility in complex agentic workflows.
Detecting AI Theory Shift via Sheaf-Theoretic Transport
This research introduces a sheaf-theoretic framework to help AI agents determine when their representational framework is no longer valid. By measuring obstruction through residual fit and constraint violations, agents can decide whether to deform existing models or extend their language.
Conditional Attribute Transformers for Steerable Sequence Generation
Conditional Attribute Transformers introduce a method to jointly estimate next-token probabilities and sequence-level attributes. This framework enables efficient steerable generation, counterfactual analysis, and per-token credit assignment without expensive sampling.
Nvidia increases salary for Jensen Huang's children
Nvidia has increased the salaries of CEO Jensen Huang's children to over $1 million annually. The company emphasized that these compensation decisions were made independently of Jensen Huang to ensure fairness.
LINE's new AI feature faces user backlash for UI clutter
LINE has introduced a new AI feature called 'Agent i' in the chat input field, which has triggered widespread complaints from users regarding accidental taps and interface clutter. Users are actively seeking ways to disable the feature due to its intrusive placement.
Honor's Lightning Robot Dominates Beijing Half-Marathon
Honor's 'Lightning' robot achieved a clean sweep of the top six positions at the Beijing Yizhuang half-marathon. The success highlights the strength of the Chinese supply chain, including key partners like GigaDevice and Rainbow Technologies.
Alibaba DAMO Academy's I2B-LPO Enhances Math Reasoning
Alibaba DAMO Academy's I2B-LPO framework, accepted at ACL 2026, addresses RLVR homogenization. It improves math reasoning accuracy by 5.3% and semantic diversity by 7.4% through optimized reasoning trajectories.
Former Kepler Robotics CEO Launches Embodied AI Startup Sota
Hu Debo, the former CEO of Kepler Robotics, has founded Sota Unlimited. The new venture shifts focus from building full-stack robots to developing the 'robot brain' for overseas markets.
SenseTime Releases SenseNova U1 Unified Multimodal Model
SenseTime has launched SenseNova U1, an open-source model featuring a native unified architecture. It integrates understanding and generation into a single representation space, moving away from traditional stitching methods.
Sony Xperia 1 VIII AI camera feature faces backlash
Sony's Xperia 1 VIII 'AI Camera Assistant' feature is under fire after promotional samples showed poor image quality compared to standard modes. Users and industry figures, including Nothing CEO Carl Pei, have questioned the marketing strategy behind the samples.
SiChuang Aerospace secures funding for rocket engine manufacturing
SiChuang Aerospace, a private Chinese manufacturer, secured millions in funding to scale its one-stop manufacturing capabilities for rocket engines, including thrust chambers and turbopumps.
Chinese Embodied AI Startup Secures Multi-Million Funding
A domestic Chinese startup specializing in the 'human learning' approach for embodied intelligence has successfully closed a funding round worth several hundred million RMB. The company aims to redefine embodied AI by prioritizing human-centric perspectives in robot learning.
Ex-DJI team launches AI-powered consumer CNC platform
Infimaker, founded by former DJI core members, secured nearly 100 million RMB to develop an easy-to-use desktop CNC. The platform integrates AI-driven CAM software and 5-axis linkage to lower the barrier for personal manufacturing.