Bilim ve Araştırma

A groundbreaking new framework called SPEX and ProxySPEX enables researchers to identify critical interactions within Large Language Models at unprecedented scale, transforming how we understand AI decision-making. This advancement moves beyond traditional explainable AI toward true mechanistic insight.

New evidence of two planets colliding in a distant solar system challenges long-held assumptions about planetary scaling and formation. The discovery, made by UW astronomers, comes after two costly experiments failed to confirm prior models.

Google DeepMind reveals multi-agent AI systems amplify errors up to 17 times, leading to costly failures. Discover the three architectural patterns separating $60M successes from canceled projects.

AlphaEvolve, once hailed as a breakthrough in AI-driven algorithm discovery, is already being superseded by autonomous frameworks like Shinka Evolve that evolve problems—not just solve them. New research reveals how AI is now inventing its own scientific questions.

Terence Tao, Fields Medalist and mathematical prodigy, has launched a new AI x Science organization to democratize scientific discovery. He predicts a future where tens of thousands of researchers will leverage AI as a collaborative partner.

Google DeepMind's Aletheia AI has achieved unprecedented autonomy in mathematical research, solving open problems without human intervention. This marks a paradigm shift from competition-level problem solving to professional discovery.

AlphaEvolve, Google DeepMind's Gemini-powered coding agent, is revolutionizing AI-driven research by co-evolving problems alongside solutions. Unlike traditional systems that optimize fixed objectives, it mirrors scientific discovery by iteratively reshaping the questions themselves.

Google AI's Groundsource leverages the Gemini model to convert millions of unstructured news articles into structured flood event data, filling critical gaps in global disaster forecasting.

Scientists have created Humanity’s Last Exam, a 2,500-question challenge designed to outpace current AI models. Results reveal even the most advanced systems struggle with specialized, expert-level knowledge — exposing a critical gap between machine performance and human cognition.

New research reveals that raising aspirations in LLM training yields unprecedented performance gains, as OpenAI integrates Promptfoo to refine AI agent behavior. Experts say this shift marks a paradigm change in how models are guided.

A landmark study reveals a critical Sim2Real gap in LLM-based user simulators, showing they overestimate agent performance by being overly cooperative and lacking human nuance. Researchers benchmarked 31 models against 451 real users.

A groundbreaking AI model developed by MIT, Mass General Brigham, and Harvard Medical School can predict which heart failure patients are likely to deteriorate within a year. The tool leverages deep learning to analyze clinical data with unprecedented accuracy.

Stanford’s Scaling Intelligence Lab has launched OpenJarvis, an open-source framework enabling fully local, on-device personal AI agents with memory, tools, and learning capabilities. The innovation marks a pivotal shift toward privacy-centric AI.

LiTo: Surface Light Field Tokenization introduces a unified 3D latent representation that jointly models geometry and view-dependent appearance, overcoming longstanding limitations in 3D reconstruction. The method leverages RGB-depth samples to encode surface light fields into compact latent vectors.

Meta's JEPA architecture has demonstrated superior performance in analyzing noisy cardiac ultrasound images, outperforming masked autoencoders and contrastive learning methods. The breakthrough stems from its non-generative, representation-prediction design.

AI engineering is no longer theoretical—it’s reshaping everyday products from medical devices to smart appliances. Drawing on insights from industry leaders, this report reveals how engineers are building robust, real-world AI systems.

Google is pioneering a novel approach to flood prediction by leveraging old news reports and large language models to transform qualitative accounts into actionable quantitative data. This innovation addresses critical data gaps in regions with limited sensor infrastructure.

Leading researchers are racing to define and engineer machine consciousness. Dr. Joscha Bach’s CIMC and the Consciousness AI project are pioneering biologically grounded approaches to create LUCID artificial systems.

A groundbreaking study challenges the assumption that LLM alignment requires diversity-seeking algorithms, finding reward-maximizing methods equally effective for moral reasoning tasks.

Google has launched Gemini Embedding 2, its first native multimodal embedding model that unifies text, image, video, and audio into a single semantic space. This breakthrough enables enterprises to process diverse data types more efficiently and at lower cost.

A groundbreaking streaming decision agent leverages partial reasoning, online replanning, and real-time adaptation to navigate dynamic environments. Drawing from AI research and agent systems, this innovation transforms autonomous decision-making in unpredictable settings.

Professor Jesse Thaler’s vision for a two-way bridge between AI and the mathematical and physical sciences is reshaping research paradigms. This synergy promises breakthroughs in both fields — and raises urgent copyright and ethical questions.

A new study by METR reveals that half of AI-generated code solutions passing the SWE-bench benchmark would be rejected by real software developers, exposing a critical gap between benchmark performance and real-world code quality.

A new study reveals that nearly half of AI-generated code solutions passing standard benchmarks are rejected by professional software engineers, exposing a gap between automated performance and real-world usability.