SourceVector-ICL: In-context Learning with Continuous Vector Representations
The authors measure each of four text encoders on a text reconstruction task (as a proxy for information preserved in embeddings) and then rank them by downstream text classification accuracy when used with vector-ICL. Across 5 classification datasets and 3 LLMs (15 configurations total), Spearman rank correlations between the two rankings are consistently positive, with 4 of 15 configurations showing correlations approaching 1. This suggests reconstruction quality is a reliable predictor of an encoder's utility as a vector-ICL input.