Auto Added by WPeMatico
What if the AI industry is optimizing for a goal that cannot be clearly defined or reliably measured? That is the central argument of a new paper by Yann LeCun, and his team, which claims that Artificial General Intelligence has become an overloaded term used in inconsistent ways across academia and industry. The research team […]
Microsoft has released Phi-4-reasoning-vision-15B, a 15 billion parameter open-weight multimodal reasoning model designed for image and text tasks that require both perception and selective reasoning. It is a compact model built to balance reasoning quality, compute efficiency, and training-data requirements, with particular strength in scientific and mathematical reasoning and understanding user interfaces. https://arxiv.org/pdf/2603.03975 What the
How can a trillion-parameter Large Language Model achieve state-of-the-art enterprise performance while simultaneously cutting its total parameter count by 33.3% and boosting pre-training efficiency by 49%? Yuan Lab AI releases Yuan3.0 Ultra, an open-source Mixture-of-Experts (MoE) large language model featuring 1T total parameters and 68.8B activated parameters. The model architecture is designed to optimize performance
Generative AI’s current trajectory relies heavily on Latent Diffusion Models (LDMs) to manage the computational cost of high-resolution synthesis. By compressing data into a lower-dimensional latent space, models can scale effectively. However, a fundamental trade-off persists: lower information density makes latents easier to learn but sacrifices reconstruction quality, while higher density enables near-perfect reconstruction but
Customizing Large Language Models (LLMs) currently presents a significant engineering trade-off between the flexibility of In-Context Learning (ICL) and the efficiency of Context Distillation (CD) or Supervised Fine-Tuning (SFT). Tokyo-based Sakana AI has proposed a new approach to bypass these constraints through cost amortization. In two of their recent papers, they introduced Text-to-LoRA (T2L) and
Perplexity has released pplx-embed, a collection of multilingual embedding models optimized for large-scale retrieval tasks. These models are designed to handle the noise and complexity of web-scale data, providing a production-ready alternative to proprietary embedding APIs. Architectural Innovations: Bidirectional Attention and Diffusion Most Large Language Models (LLMs) utilize causal, decoder-only architectures. However, for embedding tasks,
Microsoft researchers have introduced CORPGEN, an architecture-agnostic framework designed to manage the complexities of realistic organizational work through autonomous digital employees. While existing benchmarks evaluate AI agents on isolated, single tasks, real-world corporate environments require managing dozens of concurrent, interleaved tasks with complex dependencies. The research team identifies this distinct problem class as Multi-Horizon Task
In the competitive arena of Multi-Agent Reinforcement Learning (MARL), progress has long been bottlenecked by human intuition. For years, researchers have manually refined algorithms like Counterfactual Regret Minimization (CFR) and Policy Space Response Oracles (PSRO), navigating a vast combinatorial space of update rules via trial-and-error. Google DeepMind research team has now shifted this paradigm with
ByteDance Seed recently dropped a research that might change how we build reasoning AI. For years, devs and AI researchers have struggled to ‘cold-start’ Large Language Models (LLMs) into Long Chain-of-Thought (Long CoT) models. Most models lose their way or fail to transfer patterns during multi-step reasoning. The ByteDance team discovered the problem: we have
For the last few years, the AI world has followed a simple rule: if you want a Large Language Model (LLM) to solve a harder problem, make its Chain-of-Thought (CoT) longer. But new research from the University of Virginia and Google proves that ‘thinking long’ is not the same as ‘thinking hard’. The research team
