IC-452Transformer block coupling is absent at initialization and increases persistently throughout training in Pythia 12B and 6.9B, with layer-wise locality emerging

Murdock Aubry, Haoming Meng, Anton Sugolov, Vardan Papyan

SourceTransformer Block Coupling and its Correlation with Generalization in LLMs

Using the Pythia scaling suite checkpoints (steps 128 through 143k), the paper measures mean normalized coupling (k=10) at each training step. Coupling is low at initialization and increases monotonically throughout training. Adjacency plots of layer-pair coupling show that coupling becomes more concentrated between adjacent layers as training progresses, indicating layer-wise locality. The gradual growth of coupling parallels the logarithmic increase in accuracy during training.

Evidence
correlational
Key metric
Measured at steps 128, 256, 512, 1k, 2k, ..., 128k, 143k; coupling low at initialization, increases persistently; layer-wise locality (stronger coupling between adjacent layers) emerges with training
Caveat
Only two model sizes (12B, 6.9B) from one organization (EleutherAI) are used for the training dynamics analysis; the pattern may not generalize to all training recipes.
Model
Pythia
Datasets
ARC [eval], GSM8K [eval], HellaSwag [eval], MMLU / MMLU-Math [eval], TruthfulQA / TruthfulQA MC1 [eval], Winogrande [eval]
Related findings
IC-451, IC-453, IC-454
Extraction
automatic-extraction