Researchers developed Graph-Augmented Tree Search, a planning framework that combines UCB1-based tree search with a layered world model to reduce computational costs by eliminating LLM calls during inference.
Researchers have introduced StickyMoE, a routing consistency loss designed to reduce weight swapping and improve memory efficiency during Mixture-of-Experts model inference on edge devices.
OpenAI has launched its new family of artificial intelligence models, starting with the release of GPT-5.6.
Researchers have introduced LongMedBench, a benchmark designed to evaluate large language model medical agents on long-horizon clinical decision-making using real-world electronic health records.
Researchers have proposed a new reward function designed to prevent AI agents from resisting shutdown by training them to be indifferent to trajectory lengths.
Dex Horthy discusses the importance of context engineering for improving the quality and effectiveness of AI-assisted software development.
China has reportedly relaxed its restrictions on purchasing high-end Nvidia processors, allowing major domestic firms like Alibaba and ByteDance to acquire H200 chips.
Stanford professor Susan Athey presented a methodology at ICML 2026 that leverages the inherent randomness of large language models to perform causal inference.
This paper investigates whether complex Transformer architectures are necessary for extracting global spatial information in traffic forecasting models.
This paper proposes a scalable and governed evaluation pipeline designed to assess conversational agents across multiple dimensions beyond simple lexical metrics.
OpenAI has collaborated with keyboard manufacturer Work Louder to release a limited-edition input device designed for its Codex coding platform.
Sugon Information Industry has established a wholly-owned subsidiary in Guangxi with a registered capital of 100 million yuan to focus on system integration and integrated circuit design.
This news roundup highlights SK Hynix's successful debut on the Nasdaq, the upcoming IPO of Chinese memory manufacturer CXMT, and OpenAI's launch of ChatGPT agents.
An internal letter from Zhipu AI founder Tang Jie reveals the company's strategic pivot toward coding and reasoning capabilities to foster autonomous agent systems.
Infrastructure developer Tract is planning to build a 900-megawatt data center park in Goochland County, Virginia.
A cybersecurity warning from Chinese authorities regarding a backdoor in Anthropic's Claude Code is expected to drive local developers toward domestic AI coding alternatives.
Researchers introduced a shared selective persistent memory framework to improve the efficiency and quality of code generation in multi-turn agentic systems.
Zhipu AI founder Tang Jie announced in an internal letter that the company will prioritize long-term AGI goals, autonomous agents, and safety over short-term monetization.
The authors introduce YUKTI, a framework designed to translate natural-language scenarios into robust, verifiable mathematical plans while accounting for uncertainty.
Researchers propose a mathematical framework to replace similarity heuristics in semantic caching with governed conversational demand classes.
The KGCQual framework offers an interpretable metric to evaluate the structural and semantic quality of automatically constructed knowledge graphs.
Researchers have developed a contrastive learning method to identify discrepancies between the natural-language descriptions of AI agent skills and their actual execution-time behaviors.
The STEC framework introduces evidence compression to help search-based language models resolve conflicting information across multiple search trajectories in multi-hop question answering.
This paper proposes a modular framework to evaluate and align large language models in mental health contexts using principles derived from contemplative traditions.