4
New attack provides one more reason why AI browsers are a bad idea
Researchers have demonstrated a new exploit where simple logical contradictions can bypass safety guardrails in large language models integrated into web browsers.
Top stories in Research
3
Introducing GeneBench-Pro
OpenAI has introduced GeneBench-Pro, a new benchmarking tool designed to evaluate artificial intelligence performance in genomics and biological research.
3
Inside Genebench-Pro
A detailed technical overview explores the design and datasets behind OpenAI's new GeneBench-Pro evaluation framework for scientific AI.
3
ScarfBench: Benchmarking AI Agents for Enterprise Java Framework Migration
A new benchmark called ScarfBench has been introduced to evaluate the performance of artificial intelligence agents in migrating enterprise Java frameworks.
3
SkillOpt: Agent skills as trainable parameters
Microsoft researchers introduced SkillOpt, a system that treats AI agent instructions as trainable parameters to improve reliability without modifying underlying model weights.
3
Memora: A Harmonic Memory Representation Balancing Abstraction and Specificity
Microsoft Research has developed Memora, a scalable memory system designed to help AI agents retain long-term context by separating storage from retrieval.
2
Mapping Europe’s AI Workforce Opportunity
OpenAI has published a report analyzing how artificial intelligence integration could transform various job sectors and workflows across the European Union.
2
Core dump epidemiology: fixing an 18-year-old bug
OpenAI engineers utilized large-scale core dump analysis to identify and resolve a long-standing software bug and hardware fault causing infrastructure crashes.
2
Expanding our Heat Resilience data to 50+ global cities
Google Research has expanded its heat resilience data project to cover more than fifty cities globally to help combat climate change challenges.
2
Optical Scale Up Fabrics Are Limited By Manufacturing, Not Architecture
An analysis of optical scale-up fabrics suggests that manufacturing constraints, rather than architectural limitations, are the primary bottleneck for their adoption in AI clusters.
2
DiScoFormer: One transformer for density and score, across distributions
Researchers have introduced DiScoFormer, a unified transformer architecture capable of modeling both density and score functions across various distributions.
2
The Lab Mistake That Might Revolutionize Computing
Researchers are exploring new hardware materials and computing methods to reduce the high energy consumption of artificial intelligence models.
1
Accelerating Gemini Nano models on Pixel with frozen Multi-Token Prediction
Google researchers have developed a method to accelerate Gemini Nano models on Pixel devices using frozen multi-token prediction.
1
2026 BAIR Graduate Showcase
The Berkeley Artificial Intelligence Research Lab celebrated its 2026 class of doctoral graduates, highlighting their contributions to robotics, large language models, and AI safety.
1
LLMs help robots understand vague instructions and focus on key details
MIT researchers have developed a dual-language-model approach to help robots interpret vague human instructions and filter out irrelevant details.
1
How agents are transforming work
A new research paper from OpenAI details how AI agents are enhancing productivity by executing longer and more complex tasks across various job roles.
1
Understanding the brain with AI-driven explanations and experiments
Microsoft researchers introduced generative causal testing, a method that uses AI to generate and verify hypotheses about brain responses during language processing.
1
Thinking to recall: How reasoning unlocks parametric knowledge in LLMs
Google Research has explored how reasoning mechanisms can help large language models better access and utilize their stored parametric knowledge.
1
Talos: Scaling rare disease diagnosis with automated, iterative genomic reanalysis
Microsoft Research has introduced Talos, an open-source system that automates genomic reanalysis to accelerate rare disease diagnosis and reduce human review time.
1
How GPT-5 helped immunologist Derya Unutmaz solve a 3-year-old mystery
An immunologist successfully utilized OpenAI's GPT-5 Pro to resolve a three-year-old medical mystery regarding T-cell behavior.
1
ConlangCrafter Turns AI to Imagining Languages
Researchers have analyzed ConlangCrafter, an AI model designed to generate constructed languages.
1
Inaugural Music Technology Research Showcase celebrates work of new graduate program’s initial students
MIT hosted its first Music Technology Research Showcase, featuring student presentations and a keynote address on human-AI resonance in music.
1
Improving the speed and energy-efficiency of AI agents
MIT researchers have developed Murakkab, a system designed to optimize the speed and energy efficiency of multistep AI agent workflows.
1
Optimizing cloud economics with linear elastic caching
Google researchers proposed a linear elastic caching algorithm to optimize cloud computing economics.