Adaptive $k$NN graph model

Researchers have presented two new adaptive graph models, one for the k-nearest neighbors algorithm and another for whole-brain locomotion control in a simulated fruit fly, achieving real-time performance and improved interpretability.

A new adaptive graph model decouples inference latency from computational complexity in the k-nearest neighbors algorithm, a cornerstone of non-parametric classification in artificial intelligence^[1]. Existing solutions often degrade classification precision and lack adaptability, but the new model integrates a Hierarchical Navigable Small World (HNSW) graph with a pre-computed voting mechanism, achieving real-time performance without compromising classification accuracy. Meanwhile, a Whole-Brain Connectomic Graph Model enables whole-body locomotion control in a simulated fruit fly by directly instantiating the whole-brain connectome as a graph-structured neural controller^[2]. The model achieves stable performance across diverse locomotion tasks and improves sample efficiency compared to both graph and non-graph baselines. The Whole-Brain Connectomic Graph Model was first submitted on February 20, 2026, and underwent multiple revisions, with the third submission on June 14, 2026.