Multimodal in-context learning (ICL) lets large language models adapt to new tasks by conditioning on example inputs and outputs. As the paper's abstract notes, performance is tightly coupled to the choice of demonstrations, making demonstration retrieval a core challenge. The new arXiv paper introduces ActICL, a post-training method that treats demonstration selection as an agentic process rather than a fixed retrieval step.

By making the model itself responsible for choosing or managing demonstrations, ActICL aims to move beyond static retrieval pipelines. The title indicates that the approach is applied after initial training, suggesting a shift in how models acquire in-context behavior. The abstract does not disclose experimental results or implementation specifics, so the reported contribution is primarily conceptual. This is a single source, and no independent verification is available.