Pretraining Beats Architecture Benchmark Limits and the Role of Recurrent Depth Transformers in Constrained Molecular Generation

This manuscript presents a controlled benchmark study evaluating recurrent-depth transformers against pretrained language models for constrained molecular generation. Using easy, medium, and hard benchmark settings, the study demonstrates that benchmark difficulty strongly influences conclusions about architectural performance, with pretrained models outperforming architecture-first training under stringent constraints. The findings highlight the importance of robust benchmark design for evaluating AI models in molecular discovery and provide recommendations for future benchmark development. Experimental evaluations are conducted on publicly available molecular datasets, with an emphasis on reproducibility and transparent benchmark methodology.