NVIDIA introduces Polar, a framework for reinforcement learning in language agents. Polar simplifies integration with existing agent software, allowing researchers to run reinforcement learning without modifying the agent harness.
Sakana AI and University of Tokyo propose DiffusionBlocks, reducing memory usage in neural network training. Residual connections mimic Euler steps, enabling independent training of each block.
Building AI apps no longer requires complex ML knowledge. Strands Agents and AWS services enable creating intelligent agents with just 30 lines of code, simplifying AI development for AWS environments.
Stability AI unveils Stable Audio 3, featuring latent diffusion models for stereo audio generation. Models vary in size and output length, with open weights available for small and medium scales.
Amazon Quick offers a centralized observability solution for enterprise AI platforms, consolidating usage data for better tracking and analysis. By integrating with AWS services, Amazon Quick enables monitoring, analytics, and governance through a secure data lake, Amazon Athena, and Quick Sight dashboard.
Amazon Quick simplifies document creation by pulling live data from various sources and generating professional-grade documents and visuals, saving time on mechanical tasks. It supports five output types, including fully editable files that preserve formatting and data integrity, streamlining the end-to-end workflow within the Quick conversation.
AI's OSCAR addresses the challenges of INT2 KV cache quantization by using attention statistics for rotation. This method improves attention quality and reduces quantization errors, enhancing model performance significantly.
Designing a matrix inverse function using Cholesky decomposition: shorter code vs. more efficiency. Software engineering insights with AI-generated code and character design in animated films.
NVIDIA's Gated DeltaNet-2 introduces linear attention with two channel-wise gates, outperforming previous models in memory editing. Gated Delta Rule-2 separates key and value decisions, enhancing the delta-rule model's efficiency.
Perplexity's Bumblebee tool scans developer machines for vulnerable packages, extensions, and AI tool configs. It fills a gap in existing tools by checking local developer state for potential security risks.
New research by the Nous team introduces CNA, pinpointing MLP neurons responsible for refusal gates in instruct models. Ablating just 0.1% of MLP activations reduces refusal rates by over 50% without compromising output quality.
Use one-hot encoding for neural network regressors with categorical data; drop-first encoding is unnecessary and slightly less effective. Demo results show no reason to consider drop-first encoding for neural networks, confirming the advantage of one-hot encoding.
Microsoft Research's AI Frontiers lab released Fara1.5, a family of computer-use agent (CUA) models for the browser, integrated with MagenticLite. Fara1.5-27B achieves 72% task success on Online-Mind2Web, outperforming competitors like OpenAI's Operator and Google's Gemini 2.5 Computer Use.
Alibaba unveils Qwen3.7-Max, a groundbreaking AI model designed for autonomous execution of complex tasks. Qwen3.7-Max features extended-thinking mode, a 1M token context window, and impressive reasoning capabilities, setting a new standard in AI technology.
Cohere's Command A+ is an open-source MoE model optimized for high-performance agentic workflows, unifying capabilities from four prior models. It offers hardware-efficient quantization variants and significantly improves performance in agentic tasks, such as QA and coding.