The paper extracts AND/OR interaction primitives from pairs of DNNs trained for the same task and measures how many are shared across the pair. For the LLM pair LLaMA and OPT-1.3B on the SQuAD next-word task, and for the pair LLaMA and Aquila-7B, the ratio of shared interactions is substantially higher than for the BERT-base/BERT-large pair on SST-2 sentiment classification. The authors attribute this to larger models being more effective at discarding incorrectly learned interactions during training, leaving a more consistent set of task-relevant interaction primitives.
Evidence
correlational
Caveat
The specific generalization-power ratios for the LLM pairs are reported only in figures (Figures 4 and 10), not as printed numbers in the text. The BERT comparison uses the authors' fine-tuned checkpoints, so the cross-scale contrast is between released LLMs and the authors' fine-tuned BERTs.