Every story we have kept, newest first.
Looking for the daily editions? → Past editions
Page 29 of 1372
The article frames a 72-hour AI competition in which programming remains the leading capability. It argues that competition among AI coding systems is shifting from delivering code to completing end-to-end tasks.
钛媒体 · 12d ago
AI models alone do not create production value; businesses also need Harness systems that manage context, tools, state, permissions, and human handoffs. Low-cost evaluation and clear accountability often matter more than headline accuracy when deciding whether AI can enter a real workflow.
虎嗅 · 12d ago
Meigai Technology executive Yuan He argues that AI for Science will scale through narrowly defined, verifiable laboratory workflows rather than grand visions of fully autonomous factories. The hardest problems are integrating heterogeneous instruments, achieving reliable robot operation, passing acceptance tests, and turning custom projects into reusable modules.
虎嗅 · 12d ago
The CONFLICTGUI benchmark exposes execution-biased overcompliance in multimodal GUI agents when instructions conflict with themselves or the interface. CONFLICTGUARD adds feasibility verification and conditional action modulation, improving conflict-task success while preserving normal task performance.
ArXiv AI · 12d ago
Speculative Macro Commit (SMC) is a runtime system that lets a fast drafter model pre-execute multi-step tool-action chains while a larger actor model remains authoritative. It matches sequential accuracy while reducing latency by 18.59% on the τ²-Bench Telecom subset and 44.9% on AppWorld versus sequential execution.
ArXiv AI · 12d ago
This study introduces a prompt-engineering framework for personalizing general-purpose LLM/RAG teaching assistants such as Jill Watson. It combines six learner attributes with Bloom’s Taxonomy to generate 96 learner profiles and adapt responses without retraining the model.
ArXiv AI · 12d ago
The PlanFence protocol validates an agent’s pending action against the exact shared records used to create its plan. In 30 controlled workflows with post-plan revisions, PlanFence prevented every invalid action, while freshness-only validation failed every time.
ArXiv AI · 12d ago
Researchers introduce NTEP-R, a reward framework that supervises whether each tool call gathers and uses necessary evidence in agentic vision-language models. Its 8B-parameter implementation, NTEP-8B, improves search accuracy and tool-use efficiency across seven image-grounded benchmarks.
ArXiv AI · 12d ago
Researchers introduce narrative captivity, a failure mode where multi-turn LLM consultations accept an unopposed, self-justifying account as complete and align with the narrator. Across 17 LLMs and 5,078 interpersonal-conflict scenarios, judgments shifted an average of 25 percentage points from matched single-turn baselines.
ArXiv AI · 12d ago
Dude is a dual-detection multi-agent system designed to identify discrepancies between research papers and their code implementations. Its granularity-aligned negotiation and two-stage salience filtering improve recall and precision, raising F1 scores by up to 18.7% over baseline methods.
ArXiv AI · 12d ago
This arXiv study presents a governed enterprise analytics approach in which an LLM interprets questions, while deterministic policy selects and executes pre-approved analytical programs. Across the reported tests, the policy-executed analyzer fulfilled the complete answer-and-evidence contract in 110 of 110 cases, while runtime-planning systems achieved none across all datasets.
ArXiv AI · 12d ago
This research proposes Provenance Density, an interface that visualizes how many claims in a text are supported by verified evidence. In a study of 81 participants, the interface significantly improved discrimination between truthful and fabricated content, while a 200-sample audit found that consistency checks mattered more than retrieval density for dynamic queries.
ArXiv AI · 12d ago
DuplexSpeechBench-IFEval introduces 1,038 tests for measuring how full-duplex voice agents infer and follow implicit conversational instructions across eight assistant roles. Evaluation of six systems reveals that persona-only conditioning can significantly reduce floor-management adherence, while safety conflicts remain difficult to resolve.
ArXiv AI · 12d ago
AutoGraphForge is an end-to-end pipeline that generates, refutes, formalizes, and attempts to prove graph-theoretic conjectures. Its latest run produced 6,522 conjectures surviving extensive datasets and automated checks, with Lean 4 verification and neural theorem provers integrated into the workflow.
ArXiv AI · 12d ago
A new AI-driven practical English textbook uses adaptive learning to diagnose learners, generate tasks, coordinate feedback, and support teacher governance. An eight-week prototype study with 186 undergraduates improved unit completion accuracy, speaking scores, and teacher efficiency compared with a static digital textbook.
ArXiv AI · 12d ago
RIZAP disclosed that an employee mistakenly uploaded customer information to an external generative AI service used personally by the employee. The company apologized for the incident.
ITmedia AI+ (日本) · 12d ago
Economist Lan Xiaohuan discusses how China’s state-led economic model, public data infrastructure, robots, and AI relate to its economic development and competition with the United States. The interview also covers China’s record trade surplus and the need for a stronger social safety net.
SCMP Technology · 12d ago
Astro has launched Sätteri, a Rust-powered processor for Markdown and MDX. The new tool is reported to improve build speeds by up to 60%.
InfoQ中国 · 12d ago
Robots moving successfully is only the first step; the harder challenge is integrating reliable recognition into real operational workflows. Across substations, coal transport bridges, and large libraries, industrial AI must deliver accurate results continuously for years.
钛媒体 · 12d ago
A new U.S. science and technology strategy formally defines technological leadership as a national security objective.
虎嗅 · 12d ago
Kimi reportedly has annual recurring revenue of about US$300 million but is still seeking additional investment. The situation raises questions about Kimi’s capital needs, business model, and divergence from Anthropic’s development path.
钛媒体 · 12d ago
Andreessen-backed AI startup Gimlet raised $300 million in its latest funding round, reaching a $3 billion valuation. The company develops technology that divides AI tasks among different chips.
Bloomberg Technology · 12d ago
This article reviews Tencent’s year-long AI practice through the development and growth of WorkBuddy. It presents 25 practical judgments about how Tencent approached AI without simply chasing market trends.
钛媒体 · 12d ago
This article opens a historical account of China’s large-model industry and its rapid acceleration. It examines the tension between continuously scaling models and the lack of clarity about what products or applications they should ultimately become.
钛媒体 · 12d ago
Nubia NaviX Ultra, jointly developed by ZTE and ByteDance, is scheduled to launch in September as a phone centered on the Doubao assistant. It combines local AI-agent processing with ChangXin Memory’s 10667Mbps LPDDR5X, Snapdragon 8 Elite Gen 5, and a 7100mAh battery.
IT之家 · 12d ago
A RIZAP employee mistakenly uploaded customer personal and sensitive health information to an external generative AI service used privately. The exposed data reportedly included names, medical conditions, and insurance card numbers, prompting an apology from the company.
ITmedia AI+ (日本) · 12d ago
The article argues that Gemini has regained momentum, with output speed reportedly surpassing competing models and intelligence returning to the leading tier. The headline presents this as a meaningful improvement in Gemini's competitive position.
InfoQ中国 · 12d ago
The article examines how individuals can build their own DeepSeek-based coding harnesses. It questions whether developers still need to pay for subscriptions to tools such as Claude Code when comparable workflows can be assembled independently.
InfoQ中国 · 12d ago
Chaozhou Three-circle (Group) reported sharply stronger first-half earnings as AI compute infrastructure boosted demand for MLCCs. Guangdong peers Fenghua and Viiyong are also benefiting from the region’s concentration of component manufacturers.
Pandaily · 12d ago
NVIDIA has launched the beta version of Personal AI Router (PAIR), a free and open-source virtual router for local AI inference. It discovers compatible computers on the same LAN and combines their idle compute resources for AI agents and local model workloads.
cnBeta (Full RSS) · 12d ago