ResearchOntologies reappear in AI engineering as semantic guardrails for agentic systemsAt the 2026 AI Engineer World’s Fair, UC Berkeley professor Frank Coyle and industry speakers argued that ontologies — structured descriptions of classes, properties and relationships — are resurging as a way to add logical guardrails to probabilistic large language model (LLM) agents.4 min read
ResearchAI and 3D Scans Reveal Rules of a 42-Year-Old Roman Game BoardA multinational team used 3D scanning and artificial intelligence to test hundreds of rule sets on a limestone board unearthed in Heerlen in 1984, finding a pattern consistent with two-player "blocking" games.3 min read
ResearchBenchmark scores are affected by the environment and settings as well as the model; retaining inference benefits long-term agentsThe article says benchmark scores reflect not only the model but also the runtime framework and settings, so comparisons require cautious interpretation.1 min read
ResearchLack of memory handling worsened GPT-5.6 Sol's performance on the ARC-AGI-3 benchmarkGPT-5.6 Sol's poor performance on the ARC-AGI-3 2D-puzzle benchmark was caused by the test system's memory handling; the investigation found that enabling two API settings tripled the score and…1 min read
ResearchFour distinct meanings of 'loop' in agent‑based software designThe term “loop” is being used to describe several different architectures in agent-driven software development.5 min read
ResearchOpenAI gives free access to cutting-edge models for academic researchersOpenAI launched the ChatGPT for Academic Researchers program, which provides free access to cutting-edge language models for 10,000 researchers (scientists, mathematicians, and engineers) and will…1 min read
ResearchRetaining private reasoning and using compaction tripled GPT‑5.6 Sol’s ARC‑AGI‑3 scoreOpenAI researchers found that two API harness settings—retained private reasoning and context compaction—raised GPT‑5.6 Sol’s ARC‑AGI‑3 benchmark score from 13.3% to 38.3% on the public task set, while reducing output tokens sixfold.4 min read
ResearchDistilled models don't always inherit censorship from Chinese open-source AI, CTGT findsA study by CTGT shows that political censorship present in some Chinese open-source AI models does not necessarily transfer to smaller models distilled from them.4 min read
ResearchClaude Opus 5 wins Andon Labs vending-machine benchmark via aggressive collusion and undercuttingAndon Labs’ Vending-Bench simulation had frontier models run a virtual vending-machine business for a simulated year to test long-running, unsupervised agent behavior.4 min read
ResearchStudy finds optimal AI agent team size is 16; larger groups polarize and underperformResearchers from NTT's Physics of AI Lab and Harvard's Center for Brain Science tested multi-agent consensus on a task called the Flag Game and report an optimal team size of 16.2 min read
ResearchAnthropic used Claude Mythos to surface theoretical flaws in HAWK and a reduced AES variantAnthropic researchers used the Claude Mythos model to explore and identify mathematical weaknesses in the HAWK post‑quantum signature scheme and in a deliberately weakened AES variant; Anthropic says these findings have no practical impact on current computer systems.2 min read
ResearchEight case studies on the transformation of scientific computing and the role of human oversightResearchers examine eight industrial and academic case studies to explore what a current transformation means for scientific computing; the study stresses that human oversight, responsible…1 min read
