Curriculum-Inspired Progressive Example Selection for In-Context Learning

While Tailored Teaching with Balanced Difficulty (Yang et al., 2025) introduces prompt curriculum for multimodal reasoning, most active example selection strategies in ICL treat all examples equally, ignoring the pedagogical insight that learning is often most effective when examples are sequenced from simple to complex. This idea proposes a framework where the difficulty of candidate in-context examples is estimated (e.g., by model disagreement, as in active learning, or intrinsic complexity), and a curriculum is adaptively constructed—starting with easy, high-confidence cases and gradually introducing more challenging, ambiguous, or diverse examples as the model’s performance improves. Unlike previous selection methods, this approach would also re-evaluate the sequence as new data arrives, potentially reordering or swapping examples to match the model’s current state. This would be the first to systematically apply adaptive curriculum learning to the in-context example selection process in LLMs, rather than static sampling, and could yield more stable performance gains, particularly for complex or multimodal tasks.

References:

1. Tailored Teaching with Balanced Difficulty: Elevating Reasoning in Multimodal Chain-of-Thought via Prompt Curriculum. Xinglong Yang, Quan Feng, Zhongying Pan, Xiang Chen, Yu Tian, Wentong Li, Shuofei Qiao, Yuxia Geng, Xingyu Zhao, Sheng-Jun Huang (2025). arXiv.org.