LINE: LLM-based Iterative Neuron Explanations for Vision Models

Neurons learn multiple concepts simultaneously: LINE discovered that individual neurons in vision models are "polysemantic"—responding to an average of 27% more concepts than previously identified by fixed vocabulary methods—suggesting that understanding AI decision-making requires moving beyond single-concept interpretations of neural activity.

Current AI interpretation methods have blind spots: Existing neuron labeling approaches that rely on predefined concept lists miss important patterns that models actually use, which could lead to incomplete safety assessments if we're only checking for concepts we've pre-selected.

Language models can help explain vision models: By using a large language model in an iterative loop to propose and refine neuron concepts, researchers achieved state-of-the-art performance (up to 0.11 AUC improvement on ImageNet) without needing to retrain or modify the original vision model, offering a practical tool for AI auditing.

Transparency is possible without model access: LINE operates as a "black-box" method, meaning it can interpret proprietary or deployed AI systems without access to their internal workings, making it potentially applicable to real-world AI safety monitoring scenarios.

Visual explanations matter for accountability: The method produces visual explanations showing what neurons respond to, creating a more interpretable record of model behavior that could support responsible AI practices like documentation and external auditing.