Pivot-ICL: Adaptive Exemplar Selection for In-Context Learning

ACL ARR 2026 January Submission6598 Authors

Keywords: in-context learning, reasoning, large language model, adaptive treatment

Abstract:

In-context learning (ICL) enables LLMs to better complete complex tasks, but the key to success relies on selecting the right exemplars.
Prior work often selects exemplars based on similarity to the test input, but this can fail when no close match exists—especially for challenging or out-of-distribution examples.
In such cases, generic exemplars that broadly represent the task may be more helpful. This raises two key questions: (1) Which exemplars best represent the overall task distribution? (2) Which test examples are too far from available exemplars?
We propose Pivot-ICL, a method that adaptively selects either dynamic, input-specific exemplars or static, generic exemplars. We model this decision using a bipartite graph between test examples and candidate exemplars, applying well-established graph mining algorithms such as Hyperlink-Induced Topic Search (HITS) to enable weighted, iterative, and bidirectional scoring. This graph-based voting identifies both representative exemplars and which test inputs require them.Experiments show Pivot-ICL consistently improves performance across tasks, achieving a +8.8% relative gain over the best baseline.

Paper Type: Short

Research Area: Question Answering

Research Area Keywords: in-context learning, reasoning, large language model, adaptive treatment

Contribution Types: Model analysis & interpretability, NLP engineering experiment, Data analysis

Languages Studied: English

Submission Number: 6598