Investigates and prevents the loss of reasoning capabilities in language models during specialised fine-tuning. Introduces a hybrid training approach that preserves structured reasoning while adapting to domain-specific tasks.