Knowledge graph - Scientific news
Google DeepMind researchers have published a breakthrough study on "Graph--Associative Transformers" (GAT--2), a new neural architecture that natively processes knowledge graph topologies within the transformer's attention mechanism. Published on September 11, 2026, the research demonstrates that GAT--2 reduces the "context window bloat" typically associated with Large Language Models by 65% while improving reasoning accuracy on complex relational datasets. For the business sector, this breakthrough enables "Hyper--Efficient Enterprise Search," allowing LLMs to navigate massive corporate hierarchies and supply chain dependencies with significantly lower latency and computational cost.