AI Papers & Wine vol. 2 was a blast! 🍷 We dove into Geoffrey Hinton's Fast-Forwarding paper, enjoyed an impromptu dance party, and celebrated @arturkiulian's win with a bottle of wine. Thanks to all who joined us and made it an unforgettable night of AI innovation and

The paper explores the potential of AI research agents to automate the design and training of machine learning models, specifically within the context of the MLE-bench benchmark where agents compete in Kaggle competitions. It formalizes these agents as search .

The Kwai Keye-VL Technical Report presents an 8-billion-parameter multimodal foundation model designed to excel in short-video understanding while preserving general vision-language capabilities. It relies on a high-quality dataset of over 600 billion tokens a.

This paper introduces Energy-Based Transformers (EBTs), a novel class of models that enable System 2 Thinking through unsupervised learning. EBTs assign energy values to input and candidate-prediction pairs, utilizing gradient descent for optimization to enhan.

This paper investigates whether Large Language Models (LLMs) exhibit strategic intelligence in competitive settings, utilizing the Iterated Prisoner’s Dilemma (IPD) as a model for decision-making. The authors conducted evolutionary IPD tournaments with canon.

This paper discusses the evolution of multimodal reasoning, emphasizing a shift from viewing images as static contexts to utilizing them as dynamic cognitive tools within AI models. The authors introduce the 'Thinking with Images' paradigm, which is structured.

FreeMorph introduces a novel tuning-free method for image morphing, capable of generating smooth transitions between semantically diverse input images without the need for fine-tuning pre-trained diffusion models. This method overcomes the limitations of exist.

The GLM-4.1V-Thinking model, developed by Zhipu AI and Tsinghua University, is a vision-language model designed to enhance multimodal reasoning capabilities through a comprehensive training framework. The model's foundation is built on large-scale pre-training.

This paper introduces the 2-simplicial Transformer, a novel architecture that enhances the standard dot-product attention to trilinear functions, offering improved token efficiency essential for modern large language models (LLMs). By implementing an efficient.

In response to Apple's controversial paper "The Illusion of Thinking," which claimed that Large Reasoning Models (LRMs) fundamentally lack reasoning capabilities, the authors replicate and refine key benchmarks to clarify the debate. They find that LRMs strugg.

This dissertation explores the societal impact of foundation models in artificial intelligence, emphasizing their potential benefits and associated risks. It is structured around three main themes: the conceptual framing of foundation models within the economy.