This research introduces a novel approach to LLM efficiency by selectively activating only necessary parameters based on input complexity - mimicking how human brains process information.

Key Insights:

• Combines predictive coding (N400) for backbone sparsity with structural reanalysis (P600) for complex content processing
• Adapts model activation dynamically in response to cognitive load and content complexity
• Achieves significant efficiency gains while maintaining performance quality
• Bridges neuroscience principles with AI model optimization

By implementing brain-inspired mechanisms, this work demonstrates how biological intelligence principles can address critical AI efficiency bottlenecks while preserving model capabilities.

Sparse Brains are Also Adaptive Brains: Cognitive-Load-Aware Dynamic Activation for LLMs