The field of large language models is moving towards integrating symbolic knowledge with deep learning architectures to improve reasoning capabilities. Researchers are exploring various approaches to incorporate temporal logic specifications, fine-grained analysis of data synthesis pipelines, and extensions of model depth with recurrence, memory, and test-time compute scaling to enhance multi-step reasoning. A key trend is the development of neurosymbolic frameworks that provide a more structured and trustworthy alternative to traditional prompting-based methods. These frameworks leverage symbolic memory with deterministic transitions to enable robust, context-aware retrieval and transparent inference dynamics. Notable papers include: T-ILR, which proposes a neurosymbolic framework for incorporating temporal logic specifications into deep learning architectures, and FLAMES, which introduces a framework for assessing math reasoning data synthesis strategies and achieves state-of-the-art results on several benchmarks. Additionally, the extension of RetoMaton with a local, task-adaptive Weighted Finite Automaton promotes robust and interpretable reasoning, and the work on recurrence, memory, and test-time compute scaling demonstrates substantial enhancements to reasoning capabilities.