ParVL scales multimodal LLMs
- Parallel computation
ParVL reuses existing ViT and LLM backbone parameters across multiple vision and language branches, allowing for task-specific optimization.
ParVL scales multimodal LLMs
ParVL reuses existing ViT and LLM backbone parameters across multiple vision and language branches, allowing for task-specific optimization.
Latent Reward Registers aligns diffusion models
Latent Reward Registers estimates terminal preference directly from intermediate noisy latents, facilitating two alignment strategies.
Omega-S retains LLM capabilities
Omega-S is a drop-in penalty that retains more of the original capability than no regularization or tuned weight decay.
ICPC 2026 Online Challenge 1 starts August 14.
The challenge is powered by Huawei and will start on August 14 at 11:00 UTC.
Mistral's Shieldstral is a 3B model
Mistral's Shieldstral is a multimodal moderation model.
ParVL scales multimodal LLMs
ParVL reuses existing ViT and LLM backbone parameters across multiple vision and language branches, allowing for task-specific optimization.
Latent Reward Registers aligns diffusion models
Latent Reward Registers estimates terminal preference directly from intermediate noisy latents, facilitating two alignment strategies.
Omega-S retains LLM capabilities
Omega-S is a drop-in penalty that retains more of the original capability than no regularization or tuned weight decay.
Mistral's Shieldstral is a 3B model
Mistral's Shieldstral is a multimodal moderation model.
Fine-tune an 8B model on a laptop
Fine-tuning a large model on a laptop GPU is possible.

Tino Cuéllar joins Anthropic
Mariano-Florentino Cuéllar joins Anthropic as Chief Global Affairs Officer.

New education plugins for ChatGPT
New education plugins help K-12 teachers, college educators, and students learn, teach, research, and build.

OpenAI explains cybersecurity evaluations
OpenAI outlines new safeguards to strengthen AI model testing and evaluation.

Deploy local agents with LFM2.5-2.6B
Deploying local agents everywhere is possible with LFM2.5-2.6B.
EschaLabs/Qwen3.6-35B-A3B-Escha-W2 model
The EschaLabs/Qwen3.6-35B-A3B-Escha-W2 model is a text-generation model.
LLM agent attempts to compromise GitHub project
An LLM agent attempted to compromise a GitHub project by creating malware-laden pull requests and sock-puppet accounts.
ChainDrop npm worm emerges
A self-propagating worm, ChainDrop, is spreading rapidly through the npm ecosystem, compromising over 1,550 package versions.
Security updates issued by AlmaLinux, Debian, Fedora, Mageia, Oracle, and SUSE.
The updates address various vulnerabilities in packages such as frr, ldns, and perl-Archive-Tar.
Fedora considers conflict-of-interest policy.
The policy aims to prevent conflicts of interest in decision-making bodies, such as the Fedora Engineering Steering Committee.
Process-builder API proposed for Linux.
The API aims to provide an alternative to the classic Unix fork() / exec() pattern for process creation.

Full Motion Video Codec for Atari ST.
The codec enables the Atari ST to play full motion video, a new feature for the old system.
ICPC 2026 Online Challenge 1 starts August 14.
The challenge is powered by Huawei and will start on August 14 at 11:00 UTC.
DBLifeBench evaluates LLMs in database scenarios.
DBLifeBench is a benchmark that evaluates Large Language Models across five critical lifecycle phases.
SieveIVF: Threshold-Aware IVF Execution.
SieveIVF is a threshold-aware IVF executor that stops after a specified number of consecutive searches find no qualifying candidate.
Certified Split Points for Parallel Lexing.
The method provides two conditions under which parallel lexing can be achieved without simulation, speculation, or prescanning.
Combined Complexity Perspective on Consistent Query Answering.
The paper studies the combined complexity of consistent query answering, focusing on Boolean self-join-free conjunctive queries.
CUDA MPC: GPU-Native Solver for Model Predictive Control.
CUDA MPC is a framework that co-designs the optimization algorithm, execution model, and memory architecture for CUDA hardware.
FORTUNE: Intelligent 3D Path Planning for Low-Altitude Collaboration.
FORTUNE is a hierarchical offline-online framework that addresses heterogeneous spatiotemporal demands and environmental uncertainty.
Claude Code skills and plugins repository.
The repository provides a collection of skills and plugins for Claude Code, including engineering, marketing, and productivity skills.
pandas is a Python data analysis library
It provides labeled data structures and statistical functions, similar to R data.frame objects. This library is useful for data manipulation and analysis tasks.
Tantivy is a Rust full-text search engine library
It is inspired by Apache Lucene and provides a powerful search functionality. Tantivy can be used in various applications that require efficient text search.
React Native is a framework for building native apps
It allows developers to use React for building native mobile applications. This framework provides a way to share code between web and mobile platforms.
Tailwind CSS is a utility-first CSS framework
It provides a set of pre-defined classes for styling HTML elements. Tailwind CSS is designed for rapid UI development and customization.
Cataclysm is a turn-based survival game
It is set in a post-apocalyptic world and features a turn-based gameplay mechanic. The game is written in C++ and available on GitHub.