All Updates
Page 190 of 910
April 16, 2026
Aito M8 Hits 170k Sales, Leads 400k SUV Segment
Huawei's Aito M8 delivered over 170k units in first year, topping 400k RMB SUV sales. New Ruihong/Zhanlan colors with 21-inch wheels start at 369.8k RMB. Features HUAWEI ADS 4, 896-line lidar, and 705km range.
Old Android as Local AI Voice Assistant
A user repurposed an old cracked Pixel 5a into a local AI voice assistant using a llama.cpp server with gemma-3-4b model and a Termux script on the phone. The setup connects via Flask and SSH, with scrcpy for initial access. Repo enables quick setup in under 10 minutes for queries like weather.
Australian Court Slaps Warnings on Lawyer AI Use
Australia's federal court issues new rules for generative AI in proceedings amid rising AI-generated false citations. Lawyers face penalties for misleading courts with AI errors. Guidance embraces tech but stresses accountability.
$90B AI Data Centers Hit Spain
Big Tech hails northern Spain's $90 billion AI data center buildout as an EU model. Local residents experience more complex realities amid the rapid development. The story highlights both opportunities and challenges.
TRUST Agents: Multi-Agent Fake News Detector
TRUST Agents is a collaborative multi-agent framework for explainable fake news detection and fact verification. It features specialized agents for claim extraction, evidence retrieval, verification with confidence scores, and human-readable explanations. An extended version adds claim decomposition, multi-agent jury, and logic aggregation, outperforming baselines in interpretability on the LIAR benchmark.
Spatial Atlas: Compute-Grounded Spatial Reasoning
Spatial Atlas introduces compute-grounded reasoning (CGR), resolving sub-problems via deterministic computation before LM generation. It tackles FieldWorkArena for multimodal spatial QA and MLE-Bench for ML competitions using scene graphs and model routing. Achieves competitive accuracy with high interpretability.
REL Benchmark Exposes LLM Relational Limits
Researchers introduce REL, a generative benchmark framework using Relational Complexity (RC) to evaluate LLMs' higher-arity relational reasoning in algebra, chemistry, and biology. Frontier LLMs show monotonic performance drops as RC rises, persisting despite more compute or in-context learning. This highlights a core limitation in relational binding, urging benchmark reevaluation.
Platonic Reps for Permutation-Invariant Table Retrieval
This arXiv paper argues that sequential table representations in LLMs discard geometric structure, leading to brittleness under layout permutations. It introduces the Platonic Representation Hypothesis (PRH) and metrics PI/rho via CKA to quantify this bias, exposing RAG vulnerabilities. A new structure-aware encoder enforces cell header alignment for superior stability.
PhantomPolicy Detects Hidden LLM Policy Violations
LLM-based agents can commit policy-invisible violations where key facts are hidden from decision context. PhantomPolicy benchmark spans eight violation categories with 600 human-reviewed model traces from five frontier models. Sentinel enforcement framework uses counterfactual graph simulation to achieve 93% accuracy on trace-level labels.
Opinion-Aware RAG Tackles Factual Bias
Current RAG systems favor factual content, biasing against subjective opinions in real-world data like reviews. Introduces Opinion-Aware RAG with opinion extraction, entity-linked graphs, and enriched indexing to preserve opinion diversity. Experiments on e-commerce forums show +26.8% sentiment diversity and better entity/author coverage.
LLM-HYPER Solves Cold-Start Ad CTR
LLM-HYPER uses LLMs as hypernetworks to generate training-free CTR model parameters for cold-start ads via few-shot CoT prompting on multimodal content. It retrieves similar campaigns with CLIP embeddings for demonstrations. Outperforms baselines by 55.9% in NDCG@10 offline and excels in online A/B tests, now production-deployed.
EMBER: Autonomous SNN-LLM Hybrid Cognition
EMBER is a hybrid architecture placing LLMs as replaceable engines within a persistent 220k-neuron SNN substrate with STDP learning. Text embeddings are encoded via z-score top-k code for high discrimination. The SNN autonomously triggers LLM actions during idle periods without prompts, initiating first action after just 7 exchanges.
EGB Boosts Long-Horizon Tool Planning
Introduces SLATE, a large-scale benchmark for evaluating tool-augmented LLM agents in e-commerce, exposing issues in self-correction and search efficiency. Proposes Entropy-Guided Branching (EGB), an uncertainty-aware algorithm that prioritizes high-entropy decision branches. Achieves higher success rates and computational efficiency on SLATE.
AI Multi-Agent for Tumor Board Deployment
Researchers at Stanford developed a multi-agent AI system to automate patient chart summaries for Thoracic Tumor Boards, improving on manual workflows. The system was rigorously evaluated against physician gold standards using fact-based rubrics and LLM judges. It has been deployed in clinical practice with ongoing monitoring.
Aethon: Constant-Time AI Agent Instantiation
Aethon introduces a reference-based replication primitive for near-constant-time instantiation of stateful AI agents, avoiding heavy materialization. It represents agents as compositional views over shared definitions, layered memory, and local overlays. The framework uses layered inheritance and copy-on-write, enabling scalable multi-agent systems.
A-R Space Profiles Tool-Using LLM Agents
This arXiv paper introduces the A-R Behavioral Space for execution-level profiling of tool-using LLM agents. It measures Action Rate (A), Refusal Signal (R), and Divergence (D) across normative regimes and autonomy scaffolds. Results show separable behaviors shifting with reflection in risky contexts.
Gaode Secures 15 World Firsts in Embodied AI
Gaode has achieved 15 world firsts in embodied intelligence within just three months. The company plans to release its first quadruped robot soon.
Alibaba's New AI for Game Videos
Alibaba Group launched a new AI model specialized in generating videos for game development. The release supports broader efforts to commercialize AI technologies. It's aimed at enhancing game creation workflows.
Essential Energy Adopts AI for Safety
Essential Energy, an Australian utility company, is implementing AI technologies for safety inspections. The announcement highlights this strategic shift towards AI-driven operations. Additional detailed insights are teased in the article comments.
HY-World 2.0 Launches One-Click 3D Worlds
HY-World 2.0 introduces one-click generation of interactive 3D worlds from text or images. It provides editable 3D outputs compatible with Unity and Unreal Engine, including meshes, 3DGS, and point clouds. The tool features a unified model system for both synthetic and real-world scenes with real-time interactive exploration.