Скачать или смотреть Do LLMs Actually Reason? The Geometry of Attention (2512.22471)

Paper: The Bayesian Geometry of Transformer Attention (2512.22471)
Published: 27 Dec 2025.

Emergent Mind: https://www.emergentmind.com/papers/2...
arXiv: https://arxiv.org/abs/2512.22471

This video explores the groundbreaking paper "The Bayesian Geometry of Transformer Attention," which investigates whether Large Language Models are capable of exact mathematical reasoning or are simply mimicking patterns. By utilizing "Bayesian Wind Tunnels"—synthetic environments with known mathematical truths—researchers demonstrate how transformers can implement precise Bayesian inference and outperform standard neural architectures. You will see how specific internal mechanisms, like structural binding and attention-based routing, allow these models to achieve genuine algorithmic generalization on tasks far beyond their training data.