Accepted · VecDB @ VLDB 2026
Siddhesh More, Kunal JadhavArizona State University
VecDB @ VLDB 2026: 2nd Workshop on Vector Databases, 2026
Dense retrieval embeds each query as a single vector, inheriting the failure modes of whichever language model generates the representation. HyDE partially addresses this by embedding a generated hypothetical answer rather than the raw question, but the single-generator design is brittle. We propose Cross-Model Hypothesis Aggregation (CMHA), which queries N diverse LLMs from distinct model families, embeds each generated hypothesis, and uses the normalized centroid as the final query vector. Per-model errors are geometrically uncorrelated across families, so centroid averaging cancels idiosyncratic mistakes. On BEIR benchmarks with BAAI/bge-large-en-v1.5, CMHA with N=4 achieves R@10 of 0.8242 on NQ and 0.8166 on HotPotQA, recovering the 7.7pp single-model HyDE degradation and pushing 2.0pp above the direct-query baseline. We also provide, to our knowledge, the first reported empirical evidence that chain-of-thought thinking models catastrophically break HyDE-style retrieval by up to 68.2pp on HotPotQA.