Compressed Sensing for Capability Localization in Large Language Models

LLMs have modular, specialized components: The research demonstrates that large language models organize capabilities into distinct modules—removing just 5 attention heads can reduce performance on specific tasks by up to 60% while leaving other abilities largely intact, suggesting capabilities aren't distributed throughout the entire model.

Capabilities can be precisely located and manipulated: Using a compressed sensing method, researchers can identify which specific attention heads are responsible for particular skills (like math or coding) through strategic testing, opening possibilities for targeted model editing without broad performance degradation.

This principle holds across different models and sizes: The findings were validated on Llama and Qwen models ranging from 1 billion to 14 billion parameters with diverse capabilities, indicating this modular organization is a general feature of Transformer architecture rather than an anomaly in specific models.

Significant implications for AI safety and interpretability: By pinpointing which components control specific capabilities, this work enables better understanding of how LLMs work internally and creates new tools for controlling model behavior—both critical for developing safer, more trustworthy AI systems.

Practical applications for model improvement: The ability to locate and isolate capability-specific components could enable more efficient model editing, targeted fixes for problematic behaviors, and better control over which abilities models exhibit in different contexts.