📁 Frameworks

A new diffusion model, StaTS, dynamically learns the noise schedule and denoiser to improve time series forecasting. StaTS employs spectral regularization for structural preservation and a frequency-guided denoiser for enhanced reconstruction, achieving promising results with fewer sampling steps.

A new study introduces a framework for transit network design that considers demand uncertainties. The model, named 2LRC-TND, leverages machine learning and contextual stochastic optimization to model both core demand and potential service adoption, offering a more realistic alternative to fixed-demand models. The system was tested in the Atlanta metropolitan area.

The Go team has approved generic methods, reversing a longstanding position in the language's FAQ. The proposal, from Go co-designer Robert Griesemer, now moves to implementation, even as survey highlights bigger frustrations.

The new beta of the next version of Firefox lets you view two web pages side by side, with a split you can drag with your mouse.

Alibaba's team has released CoPaw, a high-performance personal workstation to help developers scale multi-channel artificial intelligence workflows. CoPaw is designed to improve memory management and the efficiency of development processes.

Introducing HumanMCP, a new large-scale dataset for evaluating the effectiveness of Model Context Protocol (MCP) servers. The dataset includes realistic, diverse, and high-quality user queries designed to simulate human interactions with 2800 tools across 308 MCP servers, addressing a gap in existing benchmarks.

NanoClaw is a smaller, security-conscious take on the OpenClaw AI agent platform. The goal is to mitigate the risks associated with unrestrained AI agents by confining execution within isolated containers.

LLmFit is a terminal tool that helps identify which LLM best fits available hardware resources. It analyzes system RAM, CPU, and GPU, evaluating models based on quality, speed, and context, suggesting the most suitable ones for execution.

Sentient launched Arena, a stress-testing environment for evaluating agentic AI in complex financial scenarios. The platform aims to improve the transparency and reliability of automation systems, enabling financial institutions to trace decision logic and ensure regulatory compliance. Franklin Templeton and Founders Fund are among the early partners.

A Reddit user expresses bewilderment regarding the popularity of OpenClaw, describing it as a wrapper with numerous pre-programmed functions. He questions whether its widespread adoption is justified, suggesting that even novice programmers could develop lighter tools tailored to their specific needs in a short amount of time.

Many organizations have deployed AI agents and automated processes, but struggle to make them collaborate efficiently and securely. The main problem is not the artificial intelligence itself, but the orchestration and coordination of these agents in complex enterprise environments.

A new study introduces AOT-SFT, a large-scale adversarial dataset, and AOT, a self-play framework to enhance the perceptual robustness of Multimodal Large Language Models (MLLMs). AOT employs a co-evolution approach between an attacker that manipulates images and a defender MLLM, forcing the latter to adapt and improve, reducing hallucinations.

Perplexity has introduced "Computer," a tool that allows users to assign complex tasks to a system of specialized AI agents. Computer breaks down the work into sub-tasks, dynamically assigning them to the most suitable models. Currently available to Perplexity Max subscribers, Computer promises to automate complex workflows.

IronCurtain is a new open source project that aims to secure and constrain AI assistant agents. The goal is to prevent unexpected or harmful behaviors that could compromise the security of data and systems.

Microsoft introduces CORPGEN, a framework for AI agents capable of managing multiple complex tasks simultaneously, simulating real-world work scenarios. CORPGEN uses hierarchical planning, isolated memories, and experiential learning to significantly improve task completion rates compared to baseline systems.

Figma has partnered with OpenAI to integrate Codex, the AI-powered coding assistant. This move follows a similar announcement regarding integration with Anthropic's Claude Code, signaling a growing interest in incorporating AI tools into design and development workflows.

OpenAI and Figma have announced a new Codex integration. The goal is to streamline the transition between visual design and code implementation, enabling teams to develop and ship products faster.

ACAR is a framework for orchestrating multiple models, using self-consistency variance to route tasks to configurations with one, two, or three models. Implemented on TEAMLLM, ACAR evaluates Claude Sonnet 4, GPT-4o, and Gemini 2.0 Flash on specific benchmarks, demonstrating 55.6% accuracy and avoiding full ensembling in 54.2% of cases. The article also highlights the limitations of retrieval augmentation and attribution based on proxy signals.

Latent Context Compilation, a new framework, addresses the challenges of deploying LLMs with long contexts. By utilizing a disposable LoRA module as a compiler, the system distills long contexts into compact, portable buffer tokens, compatible with frozen base models. This approach eliminates the need for synthetic QA pairs, preserving fine-grained details and reasoning capabilities.

A new framework, PANGAEA-GPT, leverages a hierarchical multi-agent system for the autonomous analysis of vast geoscientific data archives. The architecture, based on a Supervisor-Worker topology, enables the execution of complex workflows with minimal human intervention, improving data utilization and reusability.