Would you fine-tune an MLLM reranker or prompt one, and...

The choice depends on data availability and cost-at-scale. Prompting-based listwise reranking is ideal for early-stage development because it requires no labeled pairs and no training run. However, once you have sufficient logged relevance judgments, a fine-tuned pointwise reranker is more efficient because a small, dedicated model is significantly cheaper to serve than a large general-purpose MLLM.