Hybrid retrieval systems typically combine sparse (lexical) and dense (semantic) scores using a static interpolation weight, which fails to adapt to diverse query types. We propose a zero-overhead, pre-retrieval routing mechanism that predicts the optimal fusion weight directly from lightweight lexical query features. Our approach uses a small MLP trained end-to-end to directly optimize ranking quality, avoiding regression-based formulations. We show that this simple and efficient method matches or, in some cases, slightly exceeds thedataset-level static upper bound baseline across multiple BEIR benchmarks, while maintaining sub-millisecond latency. Our analysis further reveals that query entropy and lexical statistics are sufficient to approximate optimal fusion behavior, highlighting the inherent simplicity of the adaptive fusion problem under strict latency constraints.
Oleh Storozhev (Mon,) studied this question.