Indirect In-Context Learning via Influence-Function Demonstration Selection
A generalized in-context-learning paradigm (Indirect ICL) for selecting prompt demonstrations from a task-agnostic pool that is mostly unrelated (Mixture of Tasks) or mislabeled/poisoned (Noisy ICL). Influence functions score each candidate demonstration's effect on target-task loss and are combined with retrieval selectors (BERTScore-Recall, cosine similarity) to re-rank or reweight demonstrations, improving accuracy under task mixtures, robustness to label noise, and backdoor-attack mitigation.
B
B9
Text
UNRAVELING INDIRECT IN-CONTEXT LEARNING USING INFLUENCE FUNCTIONS Hadi Askari†, Shivanshu Gupta§
In-context learning (ICL) lets a large language model solve a task from demonstrations placed in its prompt, but its accuracy depends heavily on which demonstrations are chosen. This work defines Ind…