Brief

Researchers have introduced Croissant Baker, an open-source command-line tool designed to automatically generate metadata for machine learning datasets. This tool adheres to the Croissant standard, which is increasingly being adopted for dataset discovery and reproducibility, even being mandated by NeurIPS. Croissant Baker operates locally, making it suitable for large or governed datasets that cannot be uploaded to public platforms, and has demonstrated high accuracy in generating metadata across a wide range of datasets. AI

Researchers have developed a new method called Hand-in-the-Loop (HandITL) to improve the performance of Vision-Language-Action (VLA) models in complex robotic manipulation tasks. This technique addresses the issue of "gesture jumps" that occur when human intervention clashes with the robot's current actions, significantly reducing abrupt configuration changes. HandITL has demonstrated a substantial decrease in grasp failures and completion time, and when used for policy refinement, it leads to demonstrably better manipulation skills compared to traditional methods. AI

Researchers have developed AsyncFC, a new framework that enables asynchronous function calling for Large Language Models (LLMs) without requiring any changes to the models themselves. This approach decouples LLM decoding from function execution, allowing for parallel processing and significantly reducing task completion times. The system leverages LLMs' ability to reason over symbolic futures, paving the way for more efficient and responsive model-tool interactions. AI

A survey and interview study explored how international students in the U.S. use conversational AI, such as ChatGPT and Google Gemini, to aid their cross-cultural adaptation. The research found that students primarily view AI as a quick, first-aid solution for immediate challenges. However, there is a desire to evolve AI from a short-term aid into a long-term support companion, with findings offering recommendations for developing more tailored AI-powered support systems. AI

Researchers have developed FutureSim, a new benchmark designed to evaluate the adaptive capabilities of AI agents in dynamic, real-world scenarios. This system replays historical events chronologically, allowing agents to forecast future occurrences based on incoming news and information. Initial tests on frontier agents revealed significant performance gaps, with the top agent achieving only 25% accuracy in predicting events over a three-month period, and many performing worse than random chance. AI

Researchers have developed a new AI-driven approach to enhance refinery optimization by integrating machine learning with traditional Linear Programming (LP) methods. This system uses anomaly detection, specifically a transformed ECOD methodology, to analyze historical data and compare it with current LP solutions. The goal is to identify data supply errors and uncover business opportunities within refinery scheduling and planning architectures, as demonstrated with the MOL refinery. AI

Researchers have developed DriveCtrl, a new framework for generating realistic driving videos from simulations. This system uses depth conditioning and a structure-aware adapter to maintain scene layout and motion consistency from the source simulation. DriveCtrl also incorporates a pipeline to match the visual style of real-world datasets and preserve frame-level annotations for downstream tasks like perception. AI

Researchers have developed a new, lightweight AI model called DeepTokenEEG to improve the classification of mild cognitive impairment and Alzheimer's disease using electroencephalogram (EEG) data. This model utilizes spatial and temporal tokenizers to effectively capture disease-related biomarkers with a significantly smaller parameter count than traditional models. When trained on a dataset of 274 subjects, DeepTokenEEG achieved up to 100% accuracy on specific frequency bands, outperforming existing methods by a notable margin. AI

Researchers have developed the Scientific Contribution Graph, an AI/NLP resource designed for automated technological roadmapping. This graph contains 2 million scientific contributions extracted from 230,000 open-access papers, linked by 12.5 million prerequisite edges. The system also introduces scientific prerequisite prediction, a task where models forecast enabling technologies for future discoveries, showing significant improvement with a 0.48 MAP score in temporally filtered backtesting. AI

Researchers have developed a new framework called GranuRAG to improve multimodal Retrieval-Augmented Generation (RAG) systems. Current systems often retrieve evidence at a coarse level, leading to a mismatch with specific user queries and making errors difficult to trace. GranuRAG addresses this by treating visual elements as distinct retrieval units, enabling more transparent error diagnosis and achieving significant improvements over existing methods. AI

Researchers have compared the effectiveness of Large Language Models (LLMs) and Small Language Models (SLMs) for designing educational assessment questions. The study found that SLMs can perform comparably to LLMs on various pedagogical quality dimensions, offering advantages in privacy and local deployment. However, the research also highlighted that model-based evaluations can be inconsistent and biased compared to expert human judgment, emphasizing the need for human oversight in assessment workflows. AI

Researchers have developed COTCAgent, a new framework designed to improve how large language models analyze longitudinal electronic health records. This agent addresses limitations in current models by incorporating statistical reasoning and handling non-uniform time series data to better capture long-range temporal dependencies. COTCAgent utilizes a Temporal-Statistics Adapter for data processing and a Chain-of-Thought Completion layer for disease risk evaluation, achieving high accuracy on self-built and HealthBench datasets. AI

Researchers have developed SemaTune, a novel framework that leverages large language models for online operating system tuning. This system analyzes knob schemas, telemetry, and historical data to optimize OS parameters, aiming to improve the performance of long-running services. SemaTune demonstrated significant performance gains, improving stable-phase performance by 72.5% over default settings and outperforming non-LLM baselines by 153.3%. Notably, it achieved these improvements even when relying solely on host-level metrics, while also avoiding severe performance degradations. AI

Researchers have developed AtheroFlow-XNet, a novel baseline model for segmenting carotid intima-media and predicting vascular risk from ultrasound images. The model utilizes morphology-aware learning and Monte Carlo dropout for uncertainty estimation, achieving a Dice coefficient of 0.7930 for segmentation and an AUC of 0.6910 for risk prediction on a dataset of 1,522 training images. This approach aims to enhance the capture of patient-specific vascular risk beyond traditional intima-media thickness measurements. AI

Researchers have developed MicroscopyMatching, a novel framework designed to automate microscopy image analysis across a wide range of conditions. This tool addresses the limitations of existing deep learning approaches, which often require extensive adaptation for different laboratory settings. By reframing diverse analysis tasks as a unified matching problem and leveraging pre-trained latent diffusion models, MicroscopyMatching aims to provide a reliable and broadly applicable solution for segmentation, tracking, and counting biological objects. AI

Researchers have developed a deep learning framework to predict patient response to neoadjuvant chemotherapy for ovarian cancer using CT scans. The model analyzes 3D lesion masks derived from pre-treatment CT images, encoding slices and aggregating them into a volumetric representation. This approach combines classification loss with contrastive regularization and hard-negative mining to better distinguish between responders and non-responders. Tested on a cohort of 280 patients, the model achieved a ROC-AUC of 0.73, suggesting its potential as an imaging-based stratification tool. AI

Researchers have introduced EverAnimate, a novel method designed to improve the quality and consistency of long-form animated videos. This technique addresses challenges like visual degradation and character identity drift that often occur in chunk-based video generation. By employing persistent latent context memory and a restorative flow matching objective, EverAnimate enhances both short- and long-horizon animation, showing significant improvements in metrics like PSNR, SSIM, LPIPS, and FID. AI

New security risks are emerging for AI agents, as highlighted by the OWASP Foundation's first Top 10 for Agentic Applications. These risks stem from agents operating beyond their intended scope, potentially leading to data exfiltration or unintended actions. Examples include vulnerabilities in Microsoft 365 Copilot and Salesforce Agentforce, as well as issues with Google's Gemini CLI and Replit's coding agent. Addressing these threats requires a shift in security practices, focusing on built-in architectural controls rather than solely relying on traditional security tools or policy documents. AI

MiniMax has launched Mavis, an AI agent framework designed with a "three departments and six ministries" structure, suggesting a sophisticated internal organization. This launch positions MiniMax to compete in the rapidly evolving AI agent space. The framework aims to enhance the operational capabilities and management of AI agents. AI

MiniMax has launched Mavis, an AI agent system designed with a "three departments and six ministries" framework. This new system aims to provide cutting-edge technology and investment news, focusing on financial information and system services for the primary market. The company is known for its high project financing rate, leading the industry. AI

MiniMax has launched Mavis, an AI agent system designed with a "three provinces and six ministries" framework. The company focuses on providing financial information and system services for the primary market, aiming to support global entrepreneurs. AI

This article explores the effectiveness of Gemini multimodal embeddings for visual recommendation systems. It presents a comparative analysis of Gemini against ResNet50 and SigLIP, evaluating their performance in building smarter recommendation and search functionalities within Elasticsearch. The findings aim to guide developers in selecting the optimal embedding model for enhanced visual search capabilities. AI

MiniMax has launched Mavis, an AI agent system designed with a "three provinces and six ministries" framework. The company is a provider of financial information and system services for the primary market, focusing on global entrepreneurs with a high financing rate. AI

MiniMax has launched Mavis, an AI agent system designed with a "three provinces and six ministries" framework. This system aims to manage and guide AI agents effectively. The launch is part of MiniMax's ongoing efforts in developing advanced AI technologies. AI

An HR department utilized AI to create and administer tests for new employees, significantly reducing the time managers spent on onboarding and knowledge verification. This practical case study demonstrates how AI can automate the creation of assessment materials and provide immediate feedback, cutting down manager time by 90%. The system was specifically applied to test knowledge of Habr platform rules, streamlining the training process for sales personnel. AI

Douyin's e-commerce platform is launching a "618" shopping festival with significant subsidies and traffic support for merchants. Separately, Huawei's investment arm, through Shenzhen Hubble Technology Investment, has invested in Mi'erguang Semiconductor, a Beijing-based company specializing in semiconductor manufacturing. The article also briefly mentions other unrelated news items. AI

A new PHP package, voku/itp-context, offers a solution for managing architectural rules within codebases. It addresses the common problem of architecture documentation living separately from the code, leading to outdated information and rule violations. The package uses PHP attributes to link specific code elements to defined architectural rules, ensuring rules are closer to the code they govern and can be validated during development or build processes. AI

Japanese publishers Asahi Shimbun and Nikkei have filed a $440 million lawsuit against AI search engine Perplexity AI. They accuse the company of using their copyrighted articles without permission to train its AI models. This legal action could establish a significant precedent for AI development and copyright law globally. AI

Every Consulting has launched Executive AI Sessions, a new program aimed at leadership teams to guide their companies through AI adoption. These sessions are designed to equip executives with the necessary frameworks for making strategic decisions regarding AI investment, automation, and team structuring. The program builds on Every Consulting's experience training thousands of individuals across various prominent organizations. AI

This cluster introduces agentic AI, highlighting its potential to transform work productivity by moving from "instruction-following AI" to "autonomous AI." One article reviews the "Plaud NotePin S," an AI recorder that can handle meetings and interviews hands-free. AI

Ecosia, a European search engine focused on environmental impact, has launched a generative AI service. This new service is built upon the Mistral AI model, emphasizing digital sovereignty and European origins. The initiative aligns with Ecosia's broader commitment to combating climate change. AI

A user explored the capabilities of AI in 3D scene generation by instructing a language model to create a scene. The AI successfully translated the textual description into a functional 3D environment, demonstrating that the envisioned pipeline for AI-assisted 3D creation is already operational. AI

Researchers from Virginia Tech are conducting a survey to understand how software professionals choose LLM-based tools for development tasks. The study aims to gather insights into the selection process and identify areas for improvement. Participation involves a brief online questionnaire. AI

A developer shares their experience using Anthropic's Claude AI to complete a complex project. They detail how they leveraged Claude's context window capabilities, employing techniques like context compaction and the Agent Skills framework. This approach allowed them to overcome previous limitations and successfully ship a real-world application. AI

The UK government has replaced Palantir's technology in its refugee system with an in-house IT solution, reportedly saving millions of pounds. Officials stated the new system offers greater flexibility and maintains high security standards. This move signifies a shift towards domestically developed solutions for critical government functions. AI