Researchers propose Found-RL, a platform to enhance Reinforcement Learning (RL) in autonomous driving using foundation models. The architecture includes an asynchronous batch inference framework to overcome latency bottlenecks, diverse supervision mechanisms, and the use of CLIP for dense reward shaping. A lightweight RL model achieves near-VLM performance with real-time inference (approx. 500 FPS).
Chrome 146 beta introduces WebNN Origin Trial, paving the way for new features for neural networks directly in the browser. This update follows the release of Chrome 145, which included JPEG-XL support, and aims to further enhance the browser's capabilities.
The llama.cpp library has added support for the Kimi-K2.5 model. This integration allows users to utilize the model directly within llama.cpp, expanding the options available for local language model inference.
AMD ROCm 7.11, the open-source GPU compute stack, has been released. Concurrently, work continues on integrating ROCm packages into Ubuntu, expanding options for developers using AMD GPUs for high-performance computing workloads.
Intel today released a new version of their Compute Runtime stack and IGC graphics compiler for Level Zero and OpenCL usage with their integrated and discrete graphics. Separately they also upstreamed more SYCL code this week into mainline LLVM.
Isaac Freund's River compositor, presented at FOSDEM 2026, brings a little old-fashioned modularity and customizability to the Wayland world. This project aims to break down complex problems into smaller, more manageable parts, offering flexibility in window management.
A developer has built an open-source RAG (Retrieval-Augmented Generation) pipeline to query a dataset of over 2 million pages extracted from the "Epstein Files". The project aims to optimize semantic search and Q&A performance at scale, addressing the challenges of data cleaning, chunking, and vectorization.
A new study introduces inclusion analytics, a discourse-based framework for assessing inclusion as a dynamic process in human-AI collaborative learning. The method measures participation equity, affective climate, and epistemic equity, revealing hidden patterns in interactions.
A new study introduces Spectral Disentanglement and Enhancement (SDE), a framework aimed at improving multimodal representations. SDE separates useful signals from noise in data, optimizing alignment between feature and spectrum for more robust generalization. Results show improvements over state-of-the-art methods.
A novel approach to enhance Transformers applied to graphs, especially for graph-level tasks. Graph token serialization allows for better capture of internal dependencies and more expressive representations, overcoming the limitations of traditional single-token methods.
MCP (Multi-Control-Panel) support in llama.cpp is now available for testing. This integration introduces new features, including system message management, a CORS proxy server, and advanced tools for prompt and resource management. The goal is to provide a more comprehensive interface for interacting with models.
Plano, an open-source framework for developing AI agents, has surpassed 5000 stars on GitHub. The project focuses on small LLMs for routing and orchestration, with a framework-agnostic approach. Plano acts as a model-integrated proxy server and data plane.
Unsloth AI announced optimizations for Mixture of Experts (MoE) model training, promising 12x faster speeds and a VRAM consumption reduction of over 35%. The optimizations, based on custom Triton kernels, support architectures like gpt-oss, Qwen3, and DeepSeek, and are compatible with consumer and data center GPUs.
A user has developed a Chrome extension that uses an AI agent to automate tasks within the browser. The source code is available on GitHub, paving the way for new automation possibilities based on LLMs.
Femtobot is an agent developed in Rust, designed to operate on low-resource machines such as older Raspberry Pis or cheap VPS instances. The goal is to provide automation capabilities with a minimal footprint, avoiding the heavy dependencies typical of other stacks. It supports Telegram, local storage, and tool execution via rig-core, all in a single 10MB binary.
BiomechAgent is an AI agent that generates code for biomechanical analysis through natural language. It enables database queries, visualizations, and data interpretation without coding. A benchmark evaluates its capabilities in data retrieval, visualization, activity classification, temporal segmentation, and clinical reasoning. Biomechanically-informed instructions improve performance, but a local open-source model performs worse than a cloud-based LLM.
A Lagged Backward-Compatible Physics-Informed Neural Network (LBC-PINN) has been developed to simulate unsaturated soil consolidation under long-term loading. The framework integrates logarithmic time segmentation and transfer learning to improve accuracy and computational efficiency. Model predictions are validated against finite element method (FEM) results.
ST-Raptor is an agentic system for question answering (QA) on semi-structured tables. It combines visual editing, tree-based structural modeling, and agent-driven query resolution to improve accuracy and usability in table understanding. Experimental results show superior performance compared to existing methods.
A recent update to llama.cpp appears to improve support for the Qwen language model. This development could facilitate the execution and inference of large models on local hardware, opening new possibilities for on-premise applications and resource-constrained environments. The online discussion focuses on the potential impact of this integration.
Debian's tag2upload has finally reached general availability (GA) status, aiming to assist Debian developers and maintainers with an improved Git-based packaging workflow. The tool seeks to streamline and enhance the efficiency of software package creation and management.