Neurons Speak in Ranges: Breaking Free from Discrete Neuronal Attribution

Researchers introduce NeuronLens, a framework that interprets neural networks by analyzing activation ranges rather than individual neurons, addressing the widespread polysemanticity problem in large language models. The range-based approach enables more precise concept manipulation while minimizing unintended degradation to model performance.

The interpretability challenge in large language models stems from a fundamental architectural constraint: individual neurons encode multiple unrelated concepts simultaneously, a phenomenon called polysemanticity. This undermines traditional neuron-level attribution methods that assume one-to-one mappings between neurons and concepts. Researchers analyzing both encoder and decoder-based architectures discovered that concept-conditioned activation magnitudes cluster into distinct, often Gaussian-like distributions within individual neurons, suggesting hidden structure beneath apparent noise.

This work builds on decades of neuroscience research demonstrating that biological neurons operate within activity ranges rather than discrete on-off states. The NeuronLens framework applies this principle to artificial neural networks by mapping specific semantic concepts to activation ranges within neurons rather than assuming concept ownership of entire neurons. When researchers tested range-based interventions against traditional neuron-level masking, they achieved superior results: targeted concept manipulation succeeded while causing substantially less collateral damage to unrelated model capabilities.

For the AI development community, this represents a methodological advancement in model interpretability and control—critical challenges as language models grow increasingly sophisticated and consequential. Better interpretability tools reduce deployment risks by enabling developers to understand which concepts drive specific model behaviors. The framework potentially enables safer fine-tuning and targeted capability adjustment without the performance degradation typical of broader interventions.

Looking forward, researchers should investigate whether activation ranges generalize across different model architectures and training procedures, and whether this approach scales to multi-layer concept interactions. Broader adoption of range-based interpretation could establish new standards for responsible AI development and accountability.

• →Polysemanticity in neurons can be addressed through activation range analysis rather than discrete neuron attribution.
• →Concept-specific activation magnitudes form distinct, often Gaussian distributions within individual neurons with minimal overlap.
• →Range-based interventions achieve better concept manipulation while causing less collateral performance degradation than traditional methods.
• →This framework advances interpretability and safety measures for large language model deployment and modification.
• →The approach applies biological neuroscience principles to artificial neural network interpretation.