HexRay: An Open-Source Neuroscope for AI - Tracing Tokens, Neurons, and Decisions for Frontier AI Research, Safety, and Security
Jonathan Jaquez · 2025
We introduce HexRay, a tool for transformer model introspection that offers token-level and neuronlevel debugging capabilities, designed to support mechanistic interpretability, AI red teaming, and safety diagnostics. Built on top of TransformerLens, HexRay provides a real-time view into the inner computational pathways of transformer models, including logit-level attribution, Chain-of-Thought tracing, and activation monitoring. This paper presents HexRay's architecture, novel features, practical applications, and its potential as a safety and interpretability instrument for frontier AI systems.