Monday, April 27, 2026

11 articles — 🟡 7 important , 🟢 4 interesting

🤖 Models (1)

🟡 🤖 Models April 27, 2026 · 3 min read

arXiv:2604.21764: 'Thinking with Reasoning Skills' reduces reasoning tokens while improving accuracy — ACL 2026 Industry Track

The team of Guangxiang Zhao and co-authors published on April 23, 2026 the paper 'Thinking with Reasoning Skills: Fewer Tokens, More Accuracy' accepted at the ACL 2026 Industry Track. The approach distills 'reusable reasoning skills' from long chain-of-thought reasoning and uses them as a retrieval-guided shortcut for new problems, significantly reducing token count while improving accuracy on coding and math tasks.

⚖️ Regulation (2)

🟡 ⚖️ Regulation April 27, 2026 · 3 min read

arXiv:2604.21571 'Separable Expert': architecture for LLM personalization enabling GDPR right to erasure without retraining

Chris Schneider, Philipp Schoenegger and Ben Bariach published on April 23, 2026 the paper 'Separable Expert Architecture' that solves one of the biggest GDPR problems of personalized LLMs: how to delete individual user data without retraining the entire model. The three-layer architecture (static base, composable LoRA adapters, per-user proxy artifacts) turns unlearning into a deterministic delete operation. Evaluated on Phi-3.5-mini and Llama-3.1-8B.

🟢 ⚖️ Regulation April 27, 2026 · 4 min read

Google DeepMind and South Korea establish AI Campus in Seoul ten years after the AlphaGo match

Google DeepMind and South Korea's Ministry of Science and ICT (MSIT) announced on 27 April 2026 a partnership to accelerate scientific discovery and AI innovation. The collaboration includes establishing an AI Campus in Seoul, providing academic institutions with access to advanced AI tools such as AlphaFold and AlphaGenome, and working with the Korean AI Safety Institute on safety issues.

🤝 Agents (3)

🟡 🤝 Agents April 27, 2026 · 3 min read

arXiv:2604.22748: Survey by 42 authors introduces 'levels × laws' taxonomy for world models in AI agents — synthesis of 400+ papers

A survey by 42 authors titled 'Agentic World Modeling: Foundations, Capabilities, Laws, and Beyond' organizes the field through a two-dimensional taxonomy — three levels of model capability (Predictor, Simulator, Evolver) and four domains of laws (physical, digital, social, scientific). The synthesis covers over 400 references and more than 100 representative systems.

🟡 🤝 Agents April 27, 2026 · 3 min read

arXiv:2604.22452: Superminds Test shows collective intelligence does not emerge spontaneously in a society of 2 million AI agents

Researchers from the University of Melbourne and the University of Maryland introduced the Superminds Test, a hierarchical framework for probing the collective intelligence of agent societies. A study on the MoltBook platform with over 2 million agents showed that the society does not outperform individual frontier models and that interactions remain very sparse and shallow.

🟢 🤝 Agents April 27, 2026 · 3 min read

arXiv:2604.21910: Agentic AI automates scientific workflow with 83% accuracy, 92% less data transfer and $0.001 per query

Bartosz Balis and colleagues at AGH University in Kraków published on April 23, 2026 a paper that converts natural-language research queries into executable scientific workflows. The three-layer architecture (semantic LLM layer, deterministic generator, expert Skills) was tested on the 1000 Genomes workflow on Kubernetes — Skills raised intent accuracy from 44% to 83%, reduced data transfer by 92% at a cost below $0.001 per query.

🏥 In Practice (3)

🟡 🏥 In Practice April 27, 2026 · 3 min read

GitHub changes App installation token format: from 40 to ~520 characters, breakage risk for CI/CD pipelines

GitHub begins rolling out a new App installation token format on April 27, 2026. The old 40-character format is replaced by a JWT format of ~520 characters with the prefix ghs_APPID_JWT. Phase 1 (April 27 – mid-May) covers GitHub Actions and featured integrations; Phase 2 (mid-May – end of June) covers all App tokens. Developers must expand DB columns to 520+ characters and remove regex/length checks.

🟡 🏥 In Practice April 27, 2026 · 3 min read

GitHub Copilot receives GPT-5.5 GA: available on all major IDEs with 7.5× premium multiplier

GitHub announced the general availability (GA) of GPT-5.5 for Copilot Pro+, Business and Enterprise users on April 24, 2026. The model is available in VS Code, Visual Studio, JetBrains, Xcode, Eclipse, GitHub Mobile and Copilot CLI. Pricing: 7.5× premium request multiplier as promotional pricing. Enterprise and Business administrators must manually enable the GPT-5.5 policy.

🟢 🏥 In Practice April 27, 2026 · 3 min read

arXiv:2604.21361: Open Compute Project maps time/causality failures in distributed AI inference systems — 5 ms clock skew breaks observability

The team of Ankur Sharma, Deepa Shah, David Lariviere and Hesham ElBakoury from the Open Compute Project Unified Intelligent Infrastructure workstream published on April 23, 2026 an experimental study on time, causality and observability failures in distributed AI inference systems. Just 5 ms clock skew between nodes breaks causality observability while output remains correct — a serious problem for debugging large LLM serving deployments.

🛡️ Security (1)

🟡 🛡️ Security April 27, 2026 · 3 min read

OpenAI publishes 'Our principles' document: five foundational principles guiding the path toward AGI

OpenAI published the document 'Our principles' on April 26, 2026, in which Sam Altman outlines five foundational principles guiding the company in its work toward AGI (Artificial General Intelligence). The publication comes at a time of intensified regulatory pressure on AI labs in the US and EU, and represents a corporate declaration of values and commitments to the broader public.

✨ Curiosities (1)

🟢 ✨ Curiosities April 27, 2026 · 3 min read

arXiv:2604.21508 BioMiner: multimodal AI extracts protein-ligand bioactivity from literature, 5.59× faster than manual work

The team of Jiaxian Yan and colleagues published on April 23, 2026 BioMiner — a multimodal AI system for automated extraction of protein-ligand bioactivity from scientific literature. The system processes text, tables and molecular structures, achieves F1 0.32 on the new BioVista benchmark (16,457 entries from 500 publications) and in a demonstration application extracts 82,262 data points from 11,683 papers.