15 articles — 🟡 7 important , 🟢 8 interesting
Researchers have introduced RA-RFT, a framework that combines the retrieval of reasoning-useful examples with reinforcement fine-tuning. Unlike classic semantic search, the system selects examples that provide genuine benefit for solving the task, improving math benchmark scores by up to 7.1 points.
Researcher Il-Seok Oh has published a comprehensive tutorial positioning world modeling as the central unifying principle of modern AI. The paper distinguishes between explicit and implicit world models and describes their role in physical AI — from robotics to autonomous vehicles.
A new study demonstrates that automatically generated multi-agent systems consistently underperform simple CoT-SC, at costs up to 10 times higher. The problem lies not in the concept of multi-agent systems, but in automated design that produces architectural bloat with no functional value.
GitHub has announced the public preview of Agentic Workflows, a new system that enables coding agents within GitHub Actions to automate complex tasks — from issue triage and CI-failure analysis to vulnerability remediation and documentation updates. Personal access tokens are simultaneously eliminated from authentication.
Researchers propose a memory value function V(m) grounded in seven cognitive-psychological factors for managing the memory of long-horizon AI agents. On the blind LongMemEval benchmark it achieves a gold-evidence retention of 0.770 versus 0.657 for the uniform approach and just 0.368 for the recency method.
OrchRM is a self-supervised reward model for multi-agent orchestration that learns from intermediate execution steps without human annotations. It achieves up to 10x better token efficiency and up to 8% higher accuracy on math, web, and multi-hop tasks.
Anthropic has released minor but practical improvements to three API tools. The code execution tool now explicitly states a 90-second per-cell limit, while web_search and web_fetch gain a new response_inclusion parameter that reduces token costs in agentic pipelines.
Anthropic and DXC Technology have signed a multi-year global alliance integrating Claude into regulated industries — banking, insurance, aviation, and government. DXC's AI platform OASIS, more than 95% of whose code was generated by Claude, is already deployed with more than 50 enterprise clients.
GitHub Enterprise Server 3.21 reaches general availability with custom organization properties for automatic enterprise ruleset targeting, hierarchy view in project tables, support for more than 300 Actions jobs, improved secret scanning, and a new REST API with 24-month backward compatibility.
Nicholas Larus-Stone, Head of AI at Benchling, explains why scientific research demands an ensemble of multiple language models rather than a single one: when models agree, the data is reliable; when they diverge, something is wrong. The company holds 14 years of structured scientific data as its foundational advantage.
New research has constructed a rigorous evaluation framework for autonomous security breaches across 300 test servers at two difficulty tiers. Of 19 language models tested, none had prior knowledge of the targets — and successful breach rates reach 69.3 percent.
New research presented at ICML 2026 shows that the three most widely deployed agentic AI frameworks — LangChain, AutoGPT, and OpenAI Agents SDK — satisfy none of the derived safety principles for public-facing applications. In a simulated government service, a memory attack raised the wrongful denial rate to 88.9 percent.
CNCF project Dapr has released version 1.18 with a new verifiable execution feature — cryptographically auditable workflow history. Three new primitives allow organizations to prove which agent launched a workflow, whether execution history has been tampered with, and where a request originated in a distributed system.
