The provided research paper introduces a novel attack method, M-Attack, designed to effectively fool sophisticated commercial large vision-language models like GPT-4.5 and Gemini. The paper highlights the limitations of existing attack strategies which often produce uniform and semantically vague perturbations, failing against these robust models. M-Attack overcomes these issues by focusing on refining semantic details within localized image regions through random cropping and alignment in the embedding space, combined with a model ensemble to capture shared semantic features. This approach achieves surprisingly high success rates, exceeding 90% on several leading models, and introduces a new metric, KMRScore, for more objective evaluation of attack transferability. Ultimately, the work demonstrates a significant advancement in attacking state-of-the-art LVLMs by exploiting their reliance on detailed semantic understanding.

This paper introduces a deep learning-based method for the generalized inverse design of complex multi-port radio-frequency and sub-terahertz electromagnetic structures and integrated circuits. This innovative approach uses deep neural networks to accurately predict the properties of arbitrary-shaped designs, significantly speeding up the synthesis process compared to traditional methods. The authors demonstrate the effectiveness of this technique through various examples, including antennas, filters, and even a broadband millimeter-wave amplifier, showcasing its potential to overcome limitations of conventional design approaches and explore a broader design space. The presented methodology enables a rapid and automated synthesis of sophisticated high-frequency circuits.

Cool twitter find

R1 with vision?

The provided files detail OWL, an innovative framework built upon CAMEL-AI, designed for multi-agent collaboration to automate real-world tasks. The README.md offers a comprehensive overview of OWL's features, including toolkits for web interaction, document processing, and code execution, along with installation instructions using various methods like uv, venv, conda, and Docker. The pyproject.toml outlines the project's metadata and dependencies, while README_zh.md provides a Chinese version of the main documentation. The Python scripts run_app.py and run_app_zh.py initiate a web-based user interface in English and Chinese, respectively, for interacting with OWL.

That

Unsloth's documentation provides a comprehensive guide to fine-tuning large language models, particularly for beginners. It outlines the benefits of fine-tuning, such as improved domain knowledge and task-specific optimization, and compares it to methods like RAG. The guide details essential steps, including choosing the right model and fine-tuning method like LoRA or QLoRA, preparing datasets, understanding crucial parameters like learning rate and epochs, and avoiding overfitting or underfitting. Practical advice is given on installation, training, evaluation, running, and saving fine-tuned models using Unsloth's tools, even on limited hardware. Ultimately, the documentation aims to empower users to effectively customize LLMs for their specific needs through fine-tuning.

Updates from the PyTorch blog

This research introduces a novel, deep-learning-enabled approach for designing complex multi-port radio-frequency and sub-terahertz circuits and electromagnetic structures. The methodology overcomes limitations of traditional design methods by using deep learning models to achieve rapid synthesis of arbitrary-shaped structures with desired radiative and scattering properties. The key innovation lies in using a convolutional neural network (CNN) to predict the spectral response of electromagnetic structures, thereby eliminating the need for time-consuming electromagnetic simulations. Experimental results validate the approach through the design, fabrication, and measurement of various components, including antennas, filters, and a broadband mm-Wave amplifier. The developed method enables a new era of automated design for RF/sub-THz systems, promising enhanced performance and reduced design time.

This research addresses the challenge of improving monocular depth estimation (MDE) using unlabeled data through a novel distillation framework. The core innovation is Cross-Context Distillation, which combines local and global depth cues to enhance pseudo-label quality and model accuracy. A multi-teacher distillation approach further leverages complementary strengths of different depth estimation models for more robust predictions. The paper systematically analyzes the impact of various depth normalization strategies on pseudo-label distillation, revealing that Cross-Context Distillation significantly outperforms existing methods on benchmark datasets.Experiments validate the effectiveness of their approach, improving both fine details and overall depth consistency in MDE.