Differential Transformer V2 (DIFF V2) enhances the original model by improving inference speed, training stability, and parameterization simplicity, making it more suitable for production-level LLMs.

Sora 2 and the Sora app were developed to ensure safety in response to novel challenges from advanced video models and social creation platforms.

The article discusses ReSCALE, an adaptation of Gumbel AlphaZero MCTS that improves budget-scalability in Large Language Models for reasoning tasks without altering the model or its training.

The article introduces KidGym, a benchmark for evaluating Multimodal Large Language Models (MLLMs) using 2D grid-based tasks that assess reasoning abilities similar to those used in children’s intelligence tests.

GitHub has expanded its application security features with AI-powered detections, enhancing code safety and efficiency.

This article explains how to use Unsloth with Hugging Face Jobs for fast fine-tuning of small AI models like LiquidAI/LFM2. 2B-Instruct using coding agents, offering free credits.

The article details the authors’ attempts to prove mathematical theorems using an AI model in a recent competition.

ConsRoute is a routing framework for LLMs that improves inference efficiency by assessing semantic consistency between models at different tiers, reducing latency and cost.

The article discusses a new training strategy called ‘latent lookahead’ for transformers that allows models to ‘think ahead’ before committing to the next token, enhancing their performance on planning tasks.

The article provides a comprehensive tracker of major AI model releases, highlighting recent updates from Google, Anthropic, DeepSeek, and Zhipu AI.