LoCo Framework Localizes Concepts in Visual Autoregressive Models
October 30, 2026
LoCo is a model-agnostic framework that maps conceptual knowledge across layer, scale, and position dimensions in autoregressive models. Testing on Infinity and HunyuanImage-3.0 demonstrates precise concept localization for more controllable generative model adaptation.
HOW THIS AFFECTS YOU
●
builderThis enables more precise control and safety interventions in visual autoregressive generative pipelines.
●
researcherYou can use LoCoBench to evaluate how specific concepts are routed through next-scale generation architectures.