Transferability of multi-vector retrieval to instruction-following and reasoning tasks

Determine how well multi-vector retrieval architectures (e.g., ModernColBERT and related ColBERT-style models using MaxSim aggregation) transfer to instruction-following and reasoning-based retrieval tasks, assessing their ability to represent instruction-conditioned top-k document combinations and identifying any fundamental limitations in these settings.

Background

The paper proves and demonstrates fundamental representational limits of single-vector embedding models via connections to sign-rank. Multi-vector models (e.g., ColBERT-style approaches) are more expressive by allocating multiple token-level vectors per sequence and using MaxSim scoring, and they outperform single-vector models on the LIMIT benchmark.

However, the authors note that such models are not typically used for instruction-following or reasoning-heavy retrieval, leaving open whether their advantages carry over to tasks where queries define complex, instruction-conditioned relevance criteria.