Sharing our latest results on fine-tuning BERT-base for scientific literature classification. We achieved 94.2% accuracy on the SciPaper benchmark, improving over the baseline by 6.8%. Key findings: - Domain-adaptive pre-training on arXiv corpus improves performance significantly - Layer-wise learning rate decay is crucial for scientific text - Sentence-level contrastive learning adds 2.1% improvement Full paper submitted to EMNLP 2026. Preprint available on request.