News Around the World — 2026-09-21

Development

• InfoQ — Kubernetes 1.37 ships, graduating the Metrics API to general availability and moving Rootless Kubelet to beta, plus default-enabled HPA scale-to-zero for GPU/AI-ML workloads.

AI

• arXiv — RecreationWorld’s new computer-use agent benchmark shows the top model, GPT-6 Astra, leads with 58.1% overall accuracy but passes the full programmatic test suite on only 2.8% of its 250 tasks, exposing a gap between visual fidelity and functional correctness.
• arXiv — A parameter-free “Memory Decision Layer” for LLM agents cuts hallucination rates by about 56% under conflicting-memory conditions, by decoupling retrieved-fact confidence from consistency at ~0.14ms overhead per decision.
• arXiv — Training LLM peer-reviewers on prior AI-generated reviews causes “scientific-judgment collapse” (compressed rating distributions, reduced semantic diversity); the authors’ TrustReviewer mitigates it via curated training data and activation steering.