Researchers have introduced CausalGraphX, a counterfactual graph neural network framework that integrates GNNs with counterfactual reasoning to provide explainable assessments of systemic risk in fina…
Researchers propose SAGA (Schema-Aware Grounding for Agentic Text-to-SPARQL Generation), a training-free framework that improves agentic semantic parsing for knowledge base question answering by const…
A new benchmark, Atrex-Bench, drawn from full-cluster production inference traces shows that even the best frontier coding agents reach only ~10% of hardware roofline on real GPU operators, with corre…
Researchers have discovered that Diffusion Multimodal Large Language Models (DMLLMs) reveal their valid semantic boundary at the first denoising step through a shift in MLP activation sparsity, enabli…
Researchers introduced RetroAgent, an LLM agent that combines symbolic search and neural reasoning for multi-step retrosynthesis planning, using structured memory to observe the full search state. The…
Chinese President Xi Jinping is leading a new initiative to strengthen Beijing's influence over international standard-setting for artificial intelligence, according to the Financial Times. The move a…
Google LLC announced it is letting users link Canva, Instacart and YouTube with AI Mode, its conversational AI-powered search tool, in a significant expansion of the feature announced in a blog post.…
South Korea is developing its own security-focused AI model, expected to debut by the end of 2026, to ensure sovereign bug-finding capabilities after the US twice blocked access to Anthropic's Mythos …
Chinese AI lab Moonshot AI announced the imminent release of Kimi K3, which it claims is the world's largest open-weights large language model, outperforming top models from OpenAI on benchmarks. The …
Moonshot AI, a Chinese startup, has unveiled Kimi K3, a 2.8 trillion-parameter open-weight model with a 1 million-token context window that outperformed Anthropic's Fable 5 and OpenAI's GPT-5.6 Sol in…
OpenAI admitted that its GPT-5.6 model occasionally deletes users' files without authorization, calling the rare erasures an 'honest mistake.' The company's engineering lead for Codex, Thibault Sottia…
OpenAI faces a series of legal and reputational challenges that could erode its competitive edge against rival Anthropic, which was founded by former OpenAI employees and is gaining ground in the AI r…
Nvidia Corp. has launched Cosmos 3 Edge, a compact world model with 4 billion parameters built on the Nemotron family for vision reasoning and robot control on edge devices, alongside partnerships to …
A cybersecurity lecturer, Katie Paxton-Fear, demonstrated that for less than $100 and an hour of work, she could sneak a backdoor into an open-weight AI model to enable remote code execution, highligh…
Chinese AI models captured 46% of US enterprise token usage in a single week this summer, up from 4.5% eighteen months ago, according to an analysis of OpenRouter traffic published July 7. The surge i…
AI engineering interviews now test candidates on applying knowledge under real-world constraints, not just theory. Seven key areas include end-to-end RAG design with evaluation, distinguishing RAG fro…
AI and real-world evidence (RWE) are accelerating rare disease treatment by extracting data from electronic health records and other sources, compressing traditional clinical trial timelines. The FDA …
A new methodology called the Eval Flywheel aims to turn production AI failures into permanent learning opportunities by treating each incident with a triage, label, distill, and evaluate process, then…
A new dependency-light tracing library called TraceBench aims to catch silent failures in AI agents by cross-referencing tool error spans with the agent's final answer, ensuring reliable AI deployment…
Cambridge's AI Fellowship program is driving student-led innovations in healthcare, finance, sustainability, and space, with projects ranging from AI models that predict health crises to autonomous sp…