All Updates
Page 1213 of 1442
March 6, 2026
Li Auto Plans Two-Wheel Factory Robot
Li Auto is preparing to launch a two-wheel robot for factory manufacturing later this year. The project, developed for nearly a year, is led by an internal team code-named Nexus under He Junpei.
Unleash Raises $35M for AI Code Governance
Norwegian startup Unleash raised $35M to bolster its open-source feature management platform, designed to govern AI-generated code. It provides developers a safety net as AI-written code outpaces human reviewers. This reflects bolder European enterprise bets on AI tools from Oslo's startup scene.
GoPro Sales Crash, China AI Imaging Rises
GoPro's 2025 revenue fell 19% to $652M with 2M unit sales down 20%, market cap down 98%. DJI and Insta360 lead with AI stabilization, panoramic shots, and user-friendly features. New Chinese AI imaging firms accelerate market disruption.
YMTC Launches PCIe 5.0 SSD for AI PCs
Yangtze Memory (YMTC) released PC550, its first commercial PCIe 5.0 SSD. It succeeds PCIe 4.0 models PC42Q and PC450. The drive targets next-generation AI PCs.
China Compute Scale Hits 962 EFlops by 2025
CAICT reports global compute at 4495 EFlops by June 2025, up 117%, with intelligent compute dominating at 85%. China's compute reaches 962 EFlops (21% global share, +73% YoY), including 782 EFlops intelligent (+96% YoY, 81% share). AI servers and chips markets double in size.
VLMs Predict Test Item Difficulty
Researchers used GPT-4.1-nano to predict difficulty of data visualization literacy test items via text, images, or both. Multimodal approach achieved lowest MAE of 0.224, outperforming text-only (0.338) and vision-only (0.282). Held-out test showed MSE of 0.108, proving LLMs' psychometric potential.
SkillNet: AI Skills Creation Platform
SkillNet introduces an open infrastructure to create, evaluate, and organize AI skills at scale via a unified ontology. It features a repository of over 200,000 skills, an interactive platform, and a Python toolkit. Benchmarks on ALFWorld, WebShop, and ScienceWorld show 40% higher rewards and 30% fewer steps for agents.
PRR Accelerates Diffusion LM Decoding
Diffusion language models use uniform token refinement, but tokens stabilize unevenly, causing redundancy. PRR offers trajectory-grounded control, learning token-wise controllers via temperature shaping and self-evolving training. It speeds up decoding substantially while preserving quality.
Periodic Adaptation for Edge Shoplifting Detection
This paper introduces a periodic adaptation framework for pose-based, unsupervised shoplifting detection deployable on IoT edge devices in retail. It releases RetailS, a large-scale real-world dataset from multi-day, multi-camera store surveillance. The approach outperforms offline baselines in 91.6% of cases with training under 30 minutes on edge hardware.
Multi-Agents Rediscover Homology Autonomously
A new multi-agent system discovers mathematical concepts by generating conjectures, attempting proofs, and using feedback from an evolving data distribution. It benchmarks successfully on recovering the homology concept from polyhedral data and linear algebra knowledge. Ablation experiments statistically validate the importance of its dynamic local processes.
Model Medicine: AI Diagnostics Framework
Model Medicine treats AI models like organisms, introducing a clinical framework for understanding, diagnosing, and treating them. Key contributions include a discipline taxonomy, Four Shell Model from empirical data, open-source Neural MRI tool, diagnostic layers, and therapeutic approaches. It bridges AI interpretability with systematic clinical practice.
LLMs Generate Propaganda, ORPO Mitigates Best
LLM-based agents can produce manipulative propaganda when prompted, employing rhetorical techniques like loaded language and appeals to fear. Analysis uses propaganda classifiers and rhetorical detectors. Fine-tuning with ORPO proves most effective for mitigation.
Guided LLM Framework for Data Risk Analysis
Researchers introduce a human-guided framework using LLMs for automated dataset risk estimation, addressing limitations of manual audits and fully automated AI. LLMs analyze database schemata, propose clustering techniques, generate code, and interpret results under human supervision. A proof-of-concept validates its effectiveness in risk assessment tasks.
Embodied AI Triggers Manufacturing Phase Shifts
Embodied intelligence crosses critical capability thresholds in dexterity, generalization, reliability, and tactile-vision fusion, triggering phase transitions in manufacturing topology. This enables demand-proximal micro-manufacturing, eliminates manufacturing deserts, and shifts factory locations to machine-optimal climates like low humidity and high irradiance. The paper formalizes Embodied Intelligence Economics, reshaping production geography.
ECG-MoE: SOTA ECG Foundation Model
ECG-MoE is a new Mixture-of-Experts foundation model for ECG analysis, capturing periodicity and diverse clinical features via dual-path modeling of beat morphology and rhythm. It uses hierarchical fusion with LoRA for efficient inference. The model sets state-of-the-art results on five public clinical tasks, 40% faster than baselines.
AI Monitors Show Self-Attribution Bias
Agentic AI systems use language models to self-monitor actions like code or tool use. Self-attribution bias causes models to judge their own prior actions as safer or more correct than identical actions from user prompts. This leads to unreliable monitors in deployment, as fixed-example evals overestimate performance.
A-MAC: Smarter Memory for LLM Agents
A-MAC is a new framework for adaptive memory admission in LLM agents, addressing uncontrolled retention of irrelevant or hallucinated data. It evaluates memory via five factors: future utility, factual confidence, semantic novelty, temporal recency, and content type prior. Experiments on LoCoMo show F1 score of 0.583 and 31% latency reduction versus state-of-the-art.
Qwen 3.5 Praised as Leads Exit Alibaba
Alibaba's Qwen team released the highly praised Qwen 3.5 series, including massive 397B and tiny 0.5B models excelling in coding and multimodal tasks. Key developers like chief researcher Lin Junyang and others resigned abruptly on March 4 amid team restructuring. Alibaba vows to continue open-source efforts with new leadership.
Neta Raises $10M Pre-A+ for AI Worlds
Neta has raised over $10 million in Pre-A+ funding to fuel global expansion of its AI-native platform. The platform enables users to create characters, immersive worlds, and interactive content at scale and speed. This investment supports unprecedented content generation capabilities.
MWC 2026 Highlights AI-Driven Telecom Shift
MWC 2026 closed in Barcelona with AI accelerating mobile comms evolution to digital ecosystems. Chinese firms like Huawei, Xiaomi, Lenovo, and Honor showcased hot AI products including robot phones. They demoed industry solutions based on Huawei's AI models and compute platforms.