Hugging Face laid out the contracts for integrating models into Transformers, while several architecture papers appeared the same week.
Hugging Face explained the contracts a model must meet to be integrated into the Transformers library. The same week arXiv papers covered adaptive layer ordering, adapter training with distillation, a chess transformer simulating users, and explainability analysis.
As new models keep arriving, written integration rules affect how consistent and maintainable implementations are. Research interest is spread across structure and behavior, so library work and papers move in parallel. Note: reproducibility and adoption of the individual papers is not confirmed.
Watch for new models integrated under these contracts and for research methods such as layer reordering entering the library.