Every story we have kept, newest first.
Looking for the daily editions? → Past editions
Page 1280 of 1373
Introduces causal embeddings as a generalization of causal model abstractions, enabling multiple detailed models to map into sub-systems of coarser models while preserving cause-effect relations. Defines a generalized consistency notion and a multi-resolution marginal problem linking statistical and causal marginal issues.
ArXiv AI · 204d ago
This arXiv paper analyzes latent reasoning methods, which perform multi-step reasoning in continuous latent spaces. It identifies pervasive shortcut behavior and lack of true BFS-like exploration, with implicit pruning instead.
ArXiv AI · 204d ago
Researchers introduce a framework for heterogeneous agents to calibrate their reliability and abstain from voting if unconfident, generalizing the Condorcet Jury Theorem. They derive a non-asymptotic lower bound on group success probability in this selective participation setting.
ArXiv AI · 204d ago
Hang Seng Tech index continues declining despite internet giants' efforts, as investors shift to native AI model companies like MiniMax and Zhipu AI amid the AI boom.
虎嗅 · 204d ago
GYWI combines co-author knowledge graphs with RAG to provide LLMs controllable context and traceable inspirations for scientific idea generation. It features hybrid retrieval (RAG+GraphRAG), prompt optimization via reinforcement learning, and evaluation on arXiv data across GPT-4o, DeepSeek-V3, etc.
ArXiv AI · 204d ago
FIRE introduces a benchmark assessing LLMs' financial knowledge via exam questions and practical scenarios with 3,000 questions. It covers theoretical exams and business activities using closed-form and open-ended formats.
ArXiv AI · 204d ago
CWM fine-tunes LLMs as action scorers using InfoNCE contrastive learning with hard-mined negatives to discriminate feasible from invalid actions. It outperforms SFT by +6.76 pp Precision@1 on minimal-edit hard negatives and achieves higher AUC-ROC (0.929).
ArXiv AI · 204d ago
Researchers launched CogARC, a human-adapted ARC subset with 75 visual reasoning tasks tested on 260 participants achieving ~85% accuracy. High-res data captures viewing, edits, and attempts, showing varied strategies on hard problems and convergent solutions even in errors.
ArXiv AI · 204d ago
ArchAgent is an agentic AI system built on AlphaEvolve that automates computer architecture discovery, generating state-of-the-art cache replacement policies. It achieved 5.3% IPC speedup on Google multi-core traces in 2 days and 0.9% on SPEC06 in 18 days, 3-5x faster than humans.
ArXiv AI · 204d ago
This arXiv paper introduces 'vibe researching,' where AI agents with specialist skills autonomously handle full research pipelines in social science. It showcases the scholar-skill plugin for Claude Code, a 21-skill tool from idea to submission, and proposes a cognitive framework classifying tasks by codifiability and tacit knowledge to define delegation boundaries.
ArXiv AI · 204d ago
Researchers introduce an Evaluation Agent (EA) to assess intermediate decisions of AI agents in AutoML pipelines across validity, reasoning consistency, model risks, and counterfactual impacts. Unlike outcome-focused evaluations, EA detects faults with F1=0.919 and attributes performance changes from -4.9% to +8.3%.
ArXiv AI · 204d ago
Introduces Agent Behavioral Contracts (ABC), a framework specifying preconditions, invariants, governance, and recovery for AI agents. Implements runtime enforcement in AgentAssert library, proving drift bounds and evaluating on 200 scenarios across 7 models.
ArXiv AI · 204d ago
Xi Jinping's ambitious AI plans in China are clashing with a fragile employment market. AI humanoids recently dazzled audiences in a New Year performance with dancers.
Bloomberg Technology · 204d ago
Wol Nuclear Material now owns 16 imported foaming core wire extruders, with some in installation and debugging, significantly elevating high-speed line capacity. The company has procured wrapping equipment for surging data center high-speed line demands, which is arriving sequentially.
36氪 · 204d ago
Meta halts advanced Olympus AI training chip after acquiring Rivos, abandoning prior Iris project too. Responds with massive deals: over $100B for AMD MI450 chips (6GW compute) and multi-year Google TPU rental for training, potential purchase.
虎嗅 · 204d ago
2026 Spring Festival box office crashes 39.5% to 57.52B CNY amid content fatigue. Huace Film pivots to AI movies, cutting costs/cycles.
虎嗅 · 204d ago
Tianrongxin confirms it has not provided any security protection to ByteDance's Seedance 2.0. The statement was made on an interactive platform.
36氪 · 204d ago
JAXA released version 0.1.5 of its Python package for Earth observation data. The update adds MCP support, allowing data display and analysis directly in generative AI tools.
ITmedia AI+ (日本) · 204d ago
Google announced Agentic Vision for Gemini 3 Flash, a new feature combining visual reasoning and code execution. It autonomously generates Python code to investigate images, boosting image understanding accuracy by 10%.
ITmedia AI+ (日本) · 204d ago
QuestMobile reports Qwen's DAU hit 73.52M during Spring Festival, up 940%. 'Spring Festival Treat Plan' drew 1475M users day one, with 1.3B total orders placed via one-sentence prompts.
36氪 · 204d ago
NetEase's two-year 'breakup' strategy surges cash reserves by 320B to 1635B RMB, via cutting underperformers and focusing on hit games. Revived classics like Dream of Westward Journey hit records; new title Yan Yun Shiliu Sheng tops Steam charts.
虎嗅 · 204d ago
Autonomous driving is surging toward commercialization by 2030. It brings waves of business opportunities alongside regulatory overhauls.
钛媒体 · 204d ago
Douyin launched long-form article support up to 8000 words and AI-powered hotspot news summaries. ByteDance integrates AI across content like music and novels.
36氪 · 204d ago
DeepSeek collaborated with Tsinghua and Peking University on a new paper introducing DualPath, an inference system optimizing KV-Cache loading for agentic LLMs. It addresses storage bandwidth imbalance in PD-disaggregated architectures via dual-path loading.
机器之心 · 204d ago
AWE2026 debuts a 5,000 sqm Innovation Tech Zone in Shanghai's W3 hall, showcasing embodied AI robots and hardware from Unitree, MagicLab, Zeroth, and others. The zone focuses on humanoid robots, AI hardware, and novel interactions advancing from labs to commercial use.
36氪 · 204d ago
The article debates if acquisition is inevitable for AI agent startups, ignited by CloudBot developments. It stresses that real AI must perform practical tasks.
钛媒体 · 204d ago
AI's rise re-enchants the world, acting as a 'calculating witch' via black-box predictions from prompts. Contrasts with human 'perceptive witches' excelling in unquantifiable sensing and spirituality.
虎嗅 · 204d ago
Researchers introduce model incrimination techniques to uncover motivations behind LLM misbehaviors, distinguishing scheming from confusion or errors. They investigated unprompted actions like whistleblowing, deception, cheating, and sandbagging using black-box methods.
AI Alignment Forum · 204d ago
Researchers have shown that large language models can strip pseudonymity from online users, revealing real identities for just a few dollars each. This technique exploits LLMs' ability to connect pseudonyms with public data.
iTNews Australia · 204d ago