16 articles — 🔴 3 critical , 🟡 7 important , 🟢 6 interesting
An MIT team published MathNet, a multimodal benchmark with 30,676 olympiad math problems from 47 countries and 17 languages. Gemini-3.1-Pro achieves 78.4%, GPT-5 69.3%, and embedding models have significant difficulty finding mathematically equivalent problems.
xAI has announced that its Speech-to-Text (STT) API is moving from beta to general availability. The service supports 25 languages, offers batch and streaming modes, and is available without a waitlist — completing the voice stack alongside the previously GA-released Grok Voice Agent.
Google introduced ReasoningBank, a memory framework that enables AI agents to learn from their own successes and failures without retraining the language model. On the WebArena benchmark it achieved 8.3% higher success rate, and on SWE-Bench-Verified 4.6% with approximately 3 fewer steps per task.
OpenAI launched the Codex Labs program and strategic partnerships with Accenture, Deloitte, and KPMG to bring the Codex agent to large enterprises worldwide. The tool has reached 4 million weekly active users, offers certifications for consultants, and enterprise packages with a consumption-based billing model.
Agent-World is a new research framework from China's Renmin University that automatically generates thousands of diverse environments for training AI agents. It replaces manually crafted benchmarks with dynamic scenarios and enables evolutionary learning through co-evolution of agent and environment.
Google launched two new Deep Research agent versions in the Gemini API — deep-research-preview-04-2026 and deep-research-max-preview-04-2026 — with MCP server integration, collaborative planning, visualizations, and streaming responses. The move positions Gemini as a serious competitor to ChatGPT Deep Research and Perplexity Deep Research.
A new arXiv survey comprehensively bridges classical Multi-Agent Systems literature with the modern LLM-agent stack. The paper identifies a paradigm shift in coordination, communication protocols, and emergent behavior — from low-level state exchange to semantic reasoning.
AWS and Anthropic enable running the Claude Cowork desktop application within AWS accounts via Amazon Bedrock. Data remains under user control, models are not trained on it, and integration with IAM and CloudTrail provides enterprise-grade auditing. Payment goes through existing AWS contracts.
The STCLab SRE team uses HolmesGPT with the ReAct pattern and CNCF tools for automatic diagnosis of Kubernetes alerts. The cost is $0.04 per investigation, around 40% of alerts are resolved autonomously, and the most important lesson: quality runbooks matter more than model choice.
Researchers led by Eranga Bandara published a mobile application that locally orchestrates Gemma, Phi-3.5-mini, and Qwen2 for DSM-5 aligned psychiatric assessments. The system sends no data to the cloud and targets sensitive contexts such as the military, criminal justice, and remote healthcare.
The new DESPITE benchmark evaluated 23 language models on 12,279 robot planning tasks. Result: the best planner fails in only 0.4% of cases, but produces dangerous plans in 28.3%. Planning and safety are orthogonal capabilities — scaling models does not fix safety shortcomings.
HuggingFace published a manifesto in which Margaret Mitchell, Yacine Jernite, Clem Delangue, and 17 co-authors argue that closed AI systems are a single point of failure in cybersecurity. The text responds to Anthropic's Mythos and calls for semi-autonomous agents with auditable logs and human oversight.
GitHub enabled declarative definition of sanitizers and validators in CodeQL 2.25.2 through YAML, without writing QL code. Eight languages are supported (C/C++, C#, Go, Java/Kotlin, JS/TS, Python, Ruby, Rust), democratizing static security analysis for teams without QL experts.
