This paper proposes a simple yet effective demonstration ordering method for in-context learning (ICL), called few-shot In-Context Curriculum Learning (ICCL). Inspired by human learning processes, ICCL gradually increases the complexity of prompt demonstrations during inference. The difficulty of demonstrations can be assessed by human experts or LLM-driven metrics such as perplexity. The paper designs extensive experiments to evaluate the effectiveness of ICCL at both corpus-level and instance-level. It also investigates the formation mechanism of LLM's ICCL capability. Experimental results show that ICCL, developed during instruction-tuning, is effective for representative open-source LLMs. The code is made publicly available for further research and applications. ICCL is based on curriculum learning, which was originally introduced by Bengio et al. (2009) as a method that progressively raises the difficulty of data samples during training. Many studies have demonstrated the efficacy of curriculum learning in different models and tasks. Since instruction-tuned LLMs have shown remarkable proficiency in understanding human intentions and generating human-like text, researchers have started integrating curriculum learning during instruction tuning. However, research on the effectiveness of curriculum learning within ICL remains limited. Some methods that gradually prompt LLMs, such as Chain of Thought (CoT), have significantly enhanced the model's ability to perform complex reasoning. This inspires the application of curriculum learning for ICL. LLMs with varying performance are treated as students with varying learning abilities, and a human educator plays the role of a facilitator, guiding the learners through the curriculum. Under the human-led curriculum, the models are gradually prompted to solve complex tasks. The paper proposes ICCL, which includes two roles: curriculum constructor and curriculum learner. The curriculum constructor, which could be either human experts or LLMs, ranks the demonstrations based on their comprehension of difficulty. Subsequently, the learner is guided in progressively solving tasks. The main contributions of this paper are: (1) Proposing ICCL, a straightforward and effective demonstration ordering method, and validating its effectiveness for open-source LLMs. (2) Using perplexity as the metric to assess the difficulty of demonstration, which outperformed many superior demonstration ordering methods. (3) Comparative analysis indicates that the ICCL capability of LLMs is developed during the instruction-tuning stage. The paper evaluates ICCL on three scientific datasets: SciCite, SciNLI, and SciERC. The results show that ICCL outperforms other corpus-level methods and instance-level baselines. The performance of ICCL is stable across all LLMs and NLP tasks compared to the random baseline. The paper also explores the formation mechanism of ICCL capability, showing that it is most likely acquired during the instruction-tuning stage. The paper concludes that gradually increasing the complexity of demonstrations in prompt can achieve better performance.This paper proposes a simple yet effective demonstration ordering method for in-context learning (ICL), called few-shot In-Context Curriculum Learning (ICCL). Inspired by human learning processes, ICCL gradually increases the complexity of prompt demonstrations during inference. The difficulty of demonstrations can be assessed by human experts or LLM-driven metrics such as perplexity. The paper designs extensive experiments to evaluate the effectiveness of ICCL at both corpus-level and instance-level. It also investigates the formation mechanism of LLM's ICCL capability. Experimental results show that ICCL, developed during instruction-tuning, is effective for representative open-source LLMs. The code is made publicly available for further research and applications. ICCL is based on curriculum learning, which was originally introduced by Bengio et al. (2009) as a method that progressively raises the difficulty of data samples during training. Many studies have demonstrated the efficacy of curriculum learning in different models and tasks. Since instruction-tuned LLMs have shown remarkable proficiency in understanding human intentions and generating human-like text, researchers have started integrating curriculum learning during instruction tuning. However, research on the effectiveness of curriculum learning within ICL remains limited. Some methods that gradually prompt LLMs, such as Chain of Thought (CoT), have significantly enhanced the model's ability to perform complex reasoning. This inspires the application of curriculum learning for ICL. LLMs with varying performance are treated as students with varying learning abilities, and a human educator plays the role of a facilitator, guiding the learners through the curriculum. Under the human-led curriculum, the models are gradually prompted to solve complex tasks. The paper proposes ICCL, which includes two roles: curriculum constructor and curriculum learner. The curriculum constructor, which could be either human experts or LLMs, ranks the demonstrations based on their comprehension of difficulty. Subsequently, the learner is guided in progressively solving tasks. The main contributions of this paper are: (1) Proposing ICCL, a straightforward and effective demonstration ordering method, and validating its effectiveness for open-source LLMs. (2) Using perplexity as the metric to assess the difficulty of demonstration, which outperformed many superior demonstration ordering methods. (3) Comparative analysis indicates that the ICCL capability of LLMs is developed during the instruction-tuning stage. The paper evaluates ICCL on three scientific datasets: SciCite, SciNLI, and SciERC. The results show that ICCL outperforms other corpus-level methods and instance-level baselines. The performance of ICCL is stable across all LLMs and NLP tasks compared to the random baseline. The paper also explores the formation mechanism of ICCL capability, showing that it is most likely acquired during the instruction-tuning stage. The paper concludes that gradually increasing the complexity of demonstrations in prompt can achieve better performance.