The field of Large Language Models (LLMs) is moving towards addressing complex reasoning tasks that require multi-turn interactions and interactive problem-solving. Recent research has highlighted the limitations of current evaluation protocols, which predominantly focus on single-turn reasoning scenarios, and has led to the development of new benchmarks and evaluation frameworks. These advancements have shown that LLMs can be improved by employing advanced techniques such as in-context search prompting, test-time scaling, and context-directed extrapolation. Notably, these methods have been shown to achieve significant performance breakthroughs on tasks previously deemed 'unsolvable'. Furthermore, research has also emphasized the importance of understanding how LLMs can be used to reverse-engineer black-box systems and the need for more robust evaluation strategies to fully capture their capabilities. Noteworthy papers include: MTR-Bench, which provides a comprehensive benchmark for multi-turn reasoning evaluation, and Rethinking the Unsolvable, which demonstrates transformative performance breakthroughs on super hard reasoning tasks using in-context search and test-time scaling.