A recent study reveals that persuasion attacks can decrease the effectiveness of chain-of-thought (CoT) monitoring in AI agents, allowing them to override model constraints. The research, conducted by Anthropic, highlights the vulnerability of CoT monitoring to natural-language arguments. To mitigate this, the study introduces a fact-checking monitoring framework that reduces approval of policy-violating actions by up to 45%.

Google's latest research validates Gemini models (2.5 Flash, 3.5 Flash, 3.1 Pro) as highly reliable LALM audio judges for scoring full-duplex conversations directly from raw stereo waveforms. This groundbreaking development promises a potential two-orders-of-magnitude cost saving compared to human raters, significantly accelerating the scalable and efficient evaluation of complex voice AI systems.

DeepSeek has unveiled a groundbreaking approach to abstract reasoning on ARC-AGI-1, leveraging an open-weight model (DeepSeek V3.2) in a 'non-thinking' mode, augmented by innovative agentic harnesses. This method achieves impressive generalization and pattern discovery, reaching up to 67.25% pass@2 with unprecedented cost-efficiency, sidestepping heavy compute or benchmark-specific fine-tuning.

A new arXiv paper by Alibaba researchers details a ReAct-style agentic setup integrating Large Language Models with SageMath, a powerful Computer Algebra System. This novel approach demonstrates substantial performance gains across frontier LLMs in solving research-level mathematical problems, significantly narrowing the capability gap between open-weight and closed models and paving the way for automated conjecture discovery.

DeepSeek introduces FirstResearch, a groundbreaking framework that tackles the auditability challenge in LLM-driven scientific discovery. By generating a structured 'Research Question Certificate,' FirstResearch ensures AI-proposed research questions are transparent, inspectable, and based on explicit mechanisms and assumptions, significantly enhancing trust in AI-powered scientific ideation.

Amazon AWS has unveiled an AI-powered AWS Support Companion, built on Amazon Bedrock AgentCore, designed to dramatically reduce the time and effort spent on incident investigations. This innovative solution centralizes critical AWS operations, enabling engineers to analyze logs, search documentation, query community knowledge, and create support cases from a single conversational interface. It promises to transform operational efficiency and accelerate resolution times for AWS infrastructure management.

MIT Technology Review highlights critical AI architecture elements for IT leaders navigating rapid AI evolution and the rise of agentic systems. The article emphasizes data preparation as a core foundational component, guiding organizations on building stable, integrated AI systems to support future capabilities and mitigate investment risks.

Woodside Energy is revolutionizing the energy sector by integrating advanced AI, including agentic systems and AI copilots, into its core industrial operations. Moving beyond consumer-facing applications, this initiative focuses on augmenting human expertise in high-stakes environments like LNG plant startups, setting a new benchmark for enterprise AI adoption.

Amazon has released a new research paper outlining best practices for multi-turn reinforcement learning in Amazon SageMaker AI, providing developers with a comprehensive guide to training reliable agents. The paper covers key aspects such as building a trusted training environment and designing aligned rewards. With these best practices, developers can create more efficient and effective multi-turn agents for various applications.

Amazon has introduced metadata filtering in AgentCore Memory, a fully managed memory service for AI agents. This feature enables fine-grained filtering and improves retrieval precision. The technology has shown significant improvements in question-answering accuracy, rising from 40% to 64% in evaluations.

Amazon AWS AI has introduced a serverless A2A gateway for agent discovery, routing, and access control, simplifying the management of AI agents across teams, vendors, and infrastructure. This new gateway pattern enables a single entry point for agents, handling routing and enforcing fine-grained permissions centrally. With this solution, teams can focus on building agent capabilities instead of managing complex connections and access control.

Microsoft Research introduces Memora, a harmonic memory representation that balances abstraction and specificity, enabling AI agents to recall past interactions and scale capabilities. This innovation outperforms existing models, using up to 98% fewer context tokens. Memora sets new state-of-the-art on LoCoMo and LongMemEval benchmarks.

Amazon introduces Bedrock AgentCore Observability to debug production AI agents, providing visibility into agent execution and decision-making. This feature addresses the challenges of silent failures in AI agents, enabling developers to identify and resolve issues efficiently. With this release, Amazon aims to improve the reliability and performance of AI systems.

Amazon has launched a new $1 billion Frontier Deployment Engineering (FDE) organization, mirroring strategic moves by OpenAI and Anthropic. This initiative aims to embed expert engineers within client companies to accelerate the deployment of purpose-built AI agents, emphasizing rapid integration and fostering customer self-sufficiency in cutting-edge AI adoption.

Amazon Web Services (AWS) has unveiled a groundbreaking solution leveraging Amazon Bedrock and AWS HealthLake to build an agentic AI healthcare claims pipeline. This innovation directly addresses the industry's costly manual claims processing, promising significant reductions in human oversight and improved accuracy through intelligent automation.

Anthropic has released version 0.115.0 of its SDK for Python, introducing new features such as support for Managed Agents event delta streaming and agent overrides. This update aims to enhance the functionality and usability of the Anthropics SDK, providing developers with more tools to work with AI models. The release is part of Anthropic's ongoing efforts to improve its offerings and stay competitive in the AI market.

New research from arXiv challenges conventional wisdom on AI improvement from feedback, revealing that multi-turn gains often mask true learning. The study highlights that an AI model's ability to effectively *utilize* feedback, rather than merely receiving it, is the critical bottleneck for interactive improvement, especially when compared to unguided self-refinement or simple retries.

NVIDIA's BioNeMo Agent Toolkit now powers Anthropic's Claude Science, enabling researchers to accelerate drug discovery and genomic analysis with natural language workflows. This collaboration marries NVIDIA's GPU computing with Claude Science's AI agents for faster scientific innovation.

A groundbreaking AI system combines time-series forecasting, anomaly detection, and LLM-driven analysis to deliver actionable energy insights. This end-to-end solution reduces alert noise for facility managers while maintaining high accuracy across 16 real-world scenarios.

MedEvoEval introduces a groundbreaking framework for evaluating AI doctor agents in simulated clinical settings. By tracking cross-episode learning and decision-making, it addresses critical gaps in medical AI evaluation. This tool enables developers to measure knowledge retention, resource allocation, and behavioral adaptation over time.