Layer-wise Analysis Quantifies Positional Bias in Transformers
August 7, 2026
A layer conductance framework reveals that recency bias increases monotonically with model depth while primacy bias diminishes. These positional importance profiles are stable across diverse texts, indicating they are inherent to the transformer architecture rather than task-specific.
HOW THIS AFFECTS YOU
●
researcherYou can use these findings to understand how positional priors are encoded within specific transformer layers.