Cross-LLM Consistency in Inference: Evidence from Shared Interactions

In three linesarXiv paper demonstrating that different LLMs develop similar inference patterns despite differing architectures and training data. Advanced models share more interaction patterns when predicting the same token. Shared interactions are lower-order with weaker positive-negative cancellation than non-shared ones.

Summary generated by Claude — human-verified