Recent research highlights the challenges faced by Large Language Models (LLMs) in reasoning tasks. A new paper introduces methods to enhance LLMs' reasoning capabilities through reinforcement learning, aiming to automate the creation of high-quality reasoning data. The study benchmarks LLMs' discourse capabilities, revealing that while LLMs excel in understanding consequences, they struggle with core aspects of reasoning in realistic environments due to insufficient agent data. The proposed LEARN-BY-INTERACT method synthesizes agent data by allowing LLMs to interact with environments and adapt based on their experiences. Additionally, the paper critiques existing benchmarks for failing to accurately reflect LLMs' reasoning abilities, suggesting that current assessments may not adequately reveal their weaknesses. This research underscores the necessity for improved evaluation methods to better gauge the reasoning capabilities of LLMs.