New technique using inter-layer encoders helps LLMs make better predictions by tapping into information from hidden layers beyond just the final output.

Standard LLMs rely only on final-layer representations for predictions, but intermediate layers contain task-relevant information that could improve performance

Inter-Layer Structural Encoders (ILSE) combines information from all internal layers using a Cayley-Encoder, a geometric approach based on expander Cayley graphs

ILSE tested across 13 classification and semantic similarity tasks using 9 pre-trained models ranging from 14 million to 8 billion parameters

New approach consistently outperforms existing baselines and methods in evaluation