Every story we have kept, newest first.
Looking for the daily editions? → Past editions
Page 192 of 1375
SPOT is a model-agnostic, sampling-based framework that explains deep reinforcement learning policies through finite-horizon action trees. By recursively simulating successor states, it exposes action preferences and downstream behaviors that single-step attribution methods may miss.
ArXiv AI · 44d ago
SBCO is a self-supervised, verifier-grounded harness optimizer for planning agents with explicit constraints. It improves agent outputs through graded feedback without human labels or a self-modifying meta-agent, matching or exceeding customized baselines while using 4–5.5 times less compute.
ArXiv AI · 44d ago
ReCBM is a framework for Concept Bottleneck Models that combines semantic concept relations with uncertainty-guided refinement. Experiments indicate improved recovery under missing or incorrect concepts, support for uncertainty-aware interventions, and compact task-relevant concept selection without reducing downstream performance.
ArXiv AI · 44d ago
MIDAS is a unified framework for sentiment analysis when multimodal inputs are missing or corrupted. It disentangles shared and modality-specific information, then uses uncertainty-aware fusion to prioritize more reliable latent features.
ArXiv AI · 44d ago
MESA dynamically selects and fuses complementary memory structures for each long-horizon agent query instead of reading all memories or relying on just one. On AMA-Bench, it improves performance by 8.5% over the strongest baseline while reducing evidence tokens by 41% compared with using all structures.
ArXiv AI · 44d ago
Researchers developed a closed-loop LLM system that interprets telemetry from a 49-channel phytosensor network and directly controls agricultural hardware. In production-scale tests, it shortened cultivation cycles by 35%, reduced energy use by 18%, and discovered a lighting strategy that achieved 67.9% energy savings.
ArXiv AI · 44d ago
This research proposes using GFlowNets to train an attacker LLM that automatically probes a victim model for vulnerabilities. The approach aims to produce more creative English attacks than fixed benchmarks and introduces Turkish-language attack generation.
ArXiv AI · 44d ago
CHORUS is a post-training framework for generating high-coverage hardware verification testbench stimuli. Its consolidated 4B model reaches 88.0% Pass@1 on CVDP-ECov, beating the 671B DeepSeek-R1 by 13.5 percentage points.
ArXiv AI · 44d ago
The CASE Framework proposes four governing disciplines for enterprise agentic AI: control theory for individual agents, complex adaptive systems for collectives, supervisory cybernetics for human-agent teams, and engineering operations for fleets. Its studies report that 82% of documented production failures span multiple layers, revealing an "Emergence Gap" in current tools and practices.
ArXiv AI · 44d ago
AEROBAT is a multi-agent system that automates the end-to-end behavioral research process for AI agents, from hypothesis generation and experiment design to analysis and reporting. Across 12 target behaviors, it tested 79 hypotheses through 1,240 controlled experiments and 23,512 simulation rounds, finding moderate-to-strong evidence for 26 hypotheses.
ArXiv AI · 44d ago
Google DeepMind’s general AI safety and alignment team reportedly uses a private form to ensure that selected applicants’ resumes are seen by a human. Internal documents cited by Bloomberg say Google’s standard application system has a non-negligible chance of incorrectly filtering resumes or taking too long to route them to the team.
cnBeta (Full RSS) · 44d ago
NVIDIA has released the open model Nemotron 3.5 Lightning with 30 billion parameters. It is designed for always-on agents such as OpenClaw and reportedly delivers performance comparable to gpt-oss-120b at roughly one-quarter of the scale.
ITmedia AI+ (日本) · 44d ago
Cognition AI is reportedly in early discussions with investors for a new funding round. The financing could raise the coding startup’s valuation by more than 50%, to at least $40 billion.
Bloomberg Technology · 44d ago
極客公園盤點了 10 款新硬體，涵蓋以 Starlink 衛星定位寵物的 Fi Ultra、能自動製作自然紀錄片的 AI 假花 Petal，以及將植物數據轉化為情緒反饋的 PlantSenso。這些產品顯示，AI 正嵌入熟悉物件，並透過情緒價值、反效率設計與特定場景解決問題。
极客公园 · 44d ago
Chinese convenience-store chains are adding ovens, steamers, fryers, coffee machines, and cooking robots to counter declining foot traffic and competition from discount snack stores and instant retail. The shift depends on standardized food supply chains, automation, AI replenishment, and tighter loss management rather than simply installing kitchen equipment.
虎嗅 · 44d ago
The article explores how artificial intelligence could transform advertising and reshape the consumer decision-making process. Its central theme is whether AI can help realize greater consumer sovereignty.
钛媒体 · 44d ago
Qwen launched a broad set of features including a research mode, scheduled tasks, an office assistant, voice calls, an agent marketplace, and an open platform for third-party services. It is also testing paid tiers for office automation and usage-based pricing for video generation as Alibaba seeks to close the gap with Doubao.
虎嗅 · 44d ago
The article argues that Shenzhen and Hong Kong should deepen cooperation by preserving their distinct strengths rather than becoming identical cities. It proposes cross-border collaboration spanning university research, AI and biotech commercialization, financial services, international certification, and global expansion.
虎嗅 · 44d ago
AI agents are expanding from browser chat interfaces to desktop environments in 2026. OpenClaw and Hermes are pursuing open-source standards overseas, while WorkBuddy, Kimi Work, QwenWork, and Trae Work are competing in China.
钛媒体 · 44d ago
Medical AI is entering a phase where accuracy alone is no longer enough to distinguish products and models. The new test is whether systems can clearly explain their conclusions to clinicians and other stakeholders.
钛媒体 · 44d ago
Alibaba Cloud has reportedly reduced the delivery timeline for AI-focused data center capacity to 100 days. The shift highlights how competition in AI infrastructure is expanding beyond models and chips toward faster deployment of compute capacity.
TechNode · 44d ago
Tmall invited AI-generated star Duan Yan to appear in an advertisement, while a voice actor approached the project before production began. The episode reflects the rapid and loosely regulated growth of China’s AI short-drama industry.
钛媒体 · 44d ago
面壁智能在半年內完成超過50億融資，估值突破200億，成為端側AI領域的重要獨角獸。文章指出，其高估值仍需要更多商業化成果驗證。
钛媒体 · 44d ago
TikTok Shop在印尼市場正面臨困境，文章認為這並非單一事件，而是反映平台電商模式的結構性問題。這場變化可能影響ByteDance的市場策略與跨境電商布局。
钛媒体 · 44d ago
Suncorp is restructuring its organization to support faster and broader adoption of AI across insurance processes. The update signals a strategic shift from experimentation toward embedding AI more deeply into operational workflows.
iTNews Australia · 44d ago
IBM will deploy an NVIDIA HGX B300 cluster on IBM Cloud for Together AI's open-source model inference services under a multiyear $240 million agreement. The dedicated inference cluster is expected to enter service in Q1 2027.
IT之家 · 44d ago
iOS 27 Beta 5 adjusts the Liquid Glass transparency slider, making its highest transparency setting more see-through than before. The visual change is subtle but gives users who prefer translucent interfaces a stronger glass-like effect.
IT之家 · 44d ago
According to the article, Claude reportedly advanced work on the Riemann conjecture from 41.6% to 67.2% of a human-defined progress benchmark. The system allegedly formed a 60-member research team, assigned reviewer and plagiarism-checking roles, and drafted a paper, although the methodology and independent verification are not provided.
虎嗅 · 44d ago
Suzhou Zhicai Technology has shipped a 6-inch, high-purity MgO sputtering target qualified for MRAM magnetic tunnel junction production. The development completes China’s domestic MRAM target-material supply chain and reduces reliance on overseas suppliers.
Pandaily · 44d ago
ByteDance reportedly created a first-level department called AI Data & Security, operating alongside Seed, Flow, and Douyin. The move comes as the company is rumored to be pretraining a model that could reach 10 trillion parameters, increasing the need for centralized data acquisition, quality control, evaluation, and safety operations.
虎嗅 · 44d ago