Large Language Models (LLMs) are increasingly used as knowledge bases, but their ability to handle time-related information needs improvement. A new research paper introduces "Temporally Consistent Factuality Probing" or "TeCFaP," a way to test how well LLMs understand facts tied to specific times. Imagine asking an LLM, "What album did Linkin Park release right after Hybrid Theory?" The challenge is not just getting the correct answer ("Meteora"), but also ensuring the LLM gives the same answer regardless of how the question is phrased. This is where "temporal consistency" comes in. The research paper introduces a new dataset, TEMP-COFAC, filled with time-related queries and their paraphrased versions. Testing various LLMs, including GPT-J, Falcon, and LLaMA, revealed that they all struggle with temporal consistency. To address this, the study proposed a solution called CoTSeLF, which combines two training approaches: instruction tuning and reinforcement learning. This technique teaches LLMs to be both accurate and consistent when dealing with time-based facts. Early results show CoTSeLF significantly boosts performance. This research highlights the need to consider the time dimension when training LLMs to ensure reliable information retrieval, particularly for applications like medical diagnosis or legal analysis where accuracy and consistency are critical.