The paper "Large Language Models Can Learn Temporal Reasoning" by Siheng Xiong, Ali Payani, Ramana Kompella, and Faramarz Fekri introduces a novel framework called TG-LLM (Temporal Graph Large Language Model) to enhance the temporal reasoning (TR) capabilities of large language models (LLMs). The authors address the limitations of LLMs in handling complex temporal concepts and logic, which are crucial for tasks such as task planning and causal relation discovery. To achieve this, TG-LLM employs a two-step process: text-to-temporal graph (TG) translation and temporal graph reasoning. The first step involves converting the input text into a latent representation, specifically a temporal graph, which captures the temporal relationships and events described in the text. The second step uses Chain-of-Thought (CoT) bootstrapping and graph data augmentation to teach the LLMs to perform deliberate reasoning over the temporal graph. The authors also develop a synthetic dataset, TGQA, which is fully controllable and requires minimal supervision. This dataset is designed to fine-tune LLMs on text-to-TG translation tasks, and experiments show that the learned capabilities from this dataset can be transferred to other TR tasks and benchmarks. Key contributions of the paper include: 1. **TG-LLM Framework**: A novel paradigm for language-based TR that enhances LLMs' ability to handle temporal reasoning. 2. **Chain-of-Thought Bootstrapping**: A method to generate reliable intermediate steps for supervised fine-tuning, improving the quality of CoTs. 3. **Graph Data Augmentation**: Strategies to mitigate data insufficiency in TR tasks by introducing disturbances to the temporal graphs during training. Experiments demonstrate that TG-LLM outperforms existing methods in various temporal reasoning tasks, showing that the proposed framework significantly improves the LLMs' ability to perform complex temporal reasoning. The paper also discusses the generalizability of the learned capabilities and provides insights into future directions, such as extending the framework to more complex applications like inductive and abductive reasoning.The paper "Large Language Models Can Learn Temporal Reasoning" by Siheng Xiong, Ali Payani, Ramana Kompella, and Faramarz Fekri introduces a novel framework called TG-LLM (Temporal Graph Large Language Model) to enhance the temporal reasoning (TR) capabilities of large language models (LLMs). The authors address the limitations of LLMs in handling complex temporal concepts and logic, which are crucial for tasks such as task planning and causal relation discovery. To achieve this, TG-LLM employs a two-step process: text-to-temporal graph (TG) translation and temporal graph reasoning. The first step involves converting the input text into a latent representation, specifically a temporal graph, which captures the temporal relationships and events described in the text. The second step uses Chain-of-Thought (CoT) bootstrapping and graph data augmentation to teach the LLMs to perform deliberate reasoning over the temporal graph. The authors also develop a synthetic dataset, TGQA, which is fully controllable and requires minimal supervision. This dataset is designed to fine-tune LLMs on text-to-TG translation tasks, and experiments show that the learned capabilities from this dataset can be transferred to other TR tasks and benchmarks. Key contributions of the paper include: 1. **TG-LLM Framework**: A novel paradigm for language-based TR that enhances LLMs' ability to handle temporal reasoning. 2. **Chain-of-Thought Bootstrapping**: A method to generate reliable intermediate steps for supervised fine-tuning, improving the quality of CoTs. 3. **Graph Data Augmentation**: Strategies to mitigate data insufficiency in TR tasks by introducing disturbances to the temporal graphs during training. Experiments demonstrate that TG-LLM outperforms existing methods in various temporal reasoning tasks, showing that the proposed framework significantly improves the LLMs' ability to perform complex temporal reasoning. The paper also discusses the generalizability of the learned capabilities and provides insights into future directions, such as extending the framework to more complex applications like inductive and abductive reasoning.