Monday, May 25, 2026

8 articles — 🟡 4 important , 🟢 4 interesting

🤖 Models (1)

🟢 🤖 Models May 25, 2026 · 2 min read

Gemini API: gemini-3.1-flash-lite-preview Shut Down — Migration to GA Version Required

Google shut down the preview version of the gemini-3.1-flash-lite model on May 25, 2026, just 18 days after the stable GA (Generally Available) version was released. Development teams using the model ID gemini-3.1-flash-lite-preview need to immediately migrate to gemini-3.1-flash-lite, as the preview endpoint is no longer available.

⚖️ Regulation (1)

🟡 ⚖️ Regulation May 25, 2026 · 2 min read

UK AI Safety Institute: Agreement with Australian AISI on Joint Evaluation and Research Sharing

The British AI Safety Institute (UK AISI) and the Australian AI Safety Institute signed an agreement on May 25, 2026, on cooperation in AI evaluation best practices and research findings exchange. The partnership expands the network of international AI safety institutions and strengthens a coordinated approach to risk assessment of advanced AI systems.

🤝 Agents (1)

🟡 🤝 Agents May 25, 2026 · 2 min read

arXiv:2605.25707: AgentHijack Benchmark Reveals Critical Fragility of AI Agents in Computer Use

Researchers presented AgentHijack at ICML 2026 — a benchmark measuring the robustness of multimodal LLM agents for computer use against realistic environmental disturbances such as pop-up windows. Results show that even minor disruptions cause significant performance drops, and the authors propose the AgentHijack-Agent framework with two modules to increase robustness.

🔧 Hardware (1)

🟡 🔧 Hardware May 25, 2026 · 2 min read

AMD: Ryzen AI Max+ 395 Runs LLM Models Up to 122 Billion Parameters with Unified Memory

AMD's ROCm blog published a detailed analysis of LLM inference on the Ryzen AI Max+ 395 processor with UMA (Unified Memory Architecture). A system with 128 GB LPDDR5X memory achieves 42 tokens/s on the Qwen3.5 35B MoE model with full GPU offloading, and also supports models up to 122B parameters by combining CPU and GPU memory.

🏥 In Practice (2)

🟢 🏥 In Practice May 25, 2026 · 2 min read

arXiv:2605.25603: CIE-Scorer Detects When LLM Chain-of-Thought Reasoning Is Misrepresented

Researchers proposed CIE-Scorer, a framework that combines mechanistic interpretability with external reasoning analysis to detect unfaithful chains of thought (CoT) in LLM models. CIE-Scorer achieves state-of-the-art results on four datasets from FaithCoT-Bench with significantly lower computational costs than previous methods.

🟢 🏥 In Practice May 25, 2026 · 2 min read

arXiv:2605.25517: Study Reveals What AI Search Engines Cite — Relevance and Position Are Key

Rahul Vishwakarma, Shushant Kumar, and Ratnesh Jamidar conducted 252,000 RAG experiments across six LLMs to determine which content factors determine citation in AI search engines (GEO — Generative Engine Optimization). Topical relevance and position in the list are key; explicit pricing and a fresh date provide additional advantage, while formatting has minimal impact.

💬 Community (1)

🟢 💬 Community May 25, 2026 · 2 min read

Anthropic: Chris Olah Calls on the Church to Become a Moral Voice in AI Development

Chris Olah, co-founder of Anthropic, addressed the public with a comment on Pope Leo XIV's papal encyclical 'Magnifica humanitas' on artificial intelligence. Olah called on religious institutions to play an active role in governing AI development, as commercial pressures impede ethical decisions at AI laboratories.

🛡️ Security (1)

🟡 🛡️ Security May 25, 2026 · 2 min read

arXiv:2605.25893: D²-Monitor Dynamically Monitors Safety of Diffusion Language Models with Just 0.85M Parameters

Researchers proposed D²-Monitor, a system for dynamic safety monitoring of diffusion language models (D-LLM) that generate text via iterative denoising. D²-Monitor uses a two-stage approach based on 'safety hesitation' as a proxy for sample difficulty, achieving state-of-the-art results with fewer than 0.85 million parameters across three datasets and four D-LLM models.