Disentangle BLAZE’s aggregation effect from adapter training

Determine whether BLAZE’s cold-start performance advantage is caused by its MaxSim reranking of the same BGE-based retrieval signal used to construct the candidate pool or by adapter fine-tuning, by isolating the contribution of untrained BGE-with-MaxSim reranking.

Background

BLAZE reranks candidates with MaxSim over BGE embeddings, while the shared FAISS candidate pool is also built from BGE embeddings. Consequently, BLAZE benefits from reranking a candidate pool using the same underlying representation, unlike TRANP-CNN and COOBA, which apply independently learned scoring functions. The paper presents two non-exclusive explanations for BLAZE’s cold-start behavior—similarity-based ranking without target-specific calibration and the same-signal reranking asymmetry—but does not identify which explanation is responsible. An untrained BGE-with-MaxSim baseline is proposed as the concrete way to isolate the aggregation effect.