This research introduces Range Attribution, a novel framework that addresses the limitations of discrete neuron mapping in LLMs by recognizing that neurons operate across activation value ranges.

• Demonstrates that neurons encode different concepts at different activation ranges
• Provides a more accurate attribution of concepts to neuronal behavior
• Enables more precise and targeted manipulation of LLM outputs
• Reduces unintended interference when controlling model behavior

For security professionals, this research offers improved methods for detecting and controlling harmful outputs by manipulating specific neuronal activation ranges, rather than entire neurons, resulting in more targeted intervention with fewer side effects.

Neurons Speak in Ranges: Breaking Free from Discrete Neuronal Attribution