Transformer Block Coupling and its Correlation with Generalization in LLMs

2025-01-22 · ICLR 2025 Poster · anchor

Findings

IC-451Transformer block coupling of Jacobian singular vectors positively correlates with benchmark performance across 30+ LLMs, more strongly than parameter count, depth, or embedding dimension
IC-452Transformer block coupling is absent at initialization and increases persistently throughout training in Pythia 12B and 6.9B, with layer-wise locality emerging
IC-453Hidden representation trajectories in trained LLMs exhibit considerable linearity (mean LSS 4.25) compared to 6.54 at initialization, and linearity increases with training
IC-454Most hidden trajectories in trained LLMs exhibit exponential growth in norm as a function of depth, a property that emerges with training