The Fusion Tax: A Diagnostic Analysis of When Hybrid Retrieval Hurts — code, per-query runs, summary results, and supplementary material (v2.2, IEEE Access resubmission)
Year: 2026
Type: software
Abstract: Code, pinned environment, per-query retrieval runs, summary result files, tables, figures, and supplementary material for "The Fusion Tax: A Diagnostic Analysis of When Hybrid Retrieval Hurts" (submitted to IEEE Access, manuscript Access-2026-34033, revised version). The pipeline runs Anserini/Pyserini BM25 (official prebuilt BEIR indexes), five dense encoders (MiniLM, BGE-small/base/large, Qwen3-Embedding-0.6B), SPLADE, and three fusion functions (RRF, normalized score interpolation, CombMNZ) on seven BEIR datasets, and computes the per-query fusion tax, held-out fusion-weight selection, symmetric harm/benefit accounting by diagnostic quadrant, risk-sensitive tests, predictability and feature-importance analyses, cross-encoder reranking recovery, and efficiency measurements. Every table and figure in the article traces to a summary JSON and the script that produced it (REPRODUCE.md). The runs archive holds the per-query top-100 runs of every retriever on every corpus, from which one driver script regenerates every first-stage summary, table and figure on a CPU. Not included, because they are large and regenerable: the BEIR corpora and the document embeddings. Version 2.2 fixes a non-deterministic tie-break in the stored fused ranks (v2.0), stores every run in one canonical rank order, reruns the reranking experiment with deterministic candidates, pins model checkpoint revisions, fingerprints cached results, enforces the BEIR self-match protocol in the pipeline (a Qwen3-Embedding ArguAna run of the earlier deposits had retained the query's own document and was repaired), and adds the runs archive.
Authors: Chauhan, Aaditya, Agarwal, Arun
Language: English
Subfield: Information Systems
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7212433652