Many production RAG systems implement retrieval abstention by thresholding raw similarity scores, implicitly treating score magnitude as a confidence signal. We demonstrate that this practice degrades systematically as queries require reasoning beyond semantic matching. Across 11 retrieval architectures and 28 datasets, neural retrievers consistently assign high similarity scores to semantically related but constraint-violating documents, causing magnitude-based thresholds to collapse toward near-random abstention performance on logical and temporal reasoning tasks---a failure we term the Magnitude Mirage. To address this without computationally expensive alternatives, we conduct a large-scale empirical study of six zero-cost Query Performance Prediction (QPP) metrics across three cognitive tiers: semantic matching (BEIR), logical reasoning (BRIGHT), and temporal reasoning (TEMPO). Our central finding is that the key improvement comes from abandoning magnitude in favor of score-distribution signals: the gain from this shift exceeds the differences among distributional alternatives by a factor of 5-10$\times$. In particular, Score Gap ($s_1 - s_k$) and a practical adaptation of Score Magnitude and Variance (LSMV) improve abstention AUROC by up to 0.16 in settings where magnitude-based confidence provides little discriminative power. These methods require no additional inference, retraining, or latency, making them a practical zero-cost replacement for magnitude thresholding in deployed RAG systems.
J. Holdcroft, Abdelrahman Abdallah, Adam Jatowt· 0 citations
Reasoning-intensive temporal retrieval requires matching a query to documents whose relevance depends on shared temporal reasoning rather than lexical overlap. Expanding a query into several reformulations that make its temporal intent explicit, and retrieving with each, supplies this reasoning, but fusing the resulting rankings with equal weights wastes accuracy: for any single query, only some reformulations are reliable. We propose query-difficulty-gated fusion of reasoning views. From each view we read an eight-dimensional signature of its score distribution, built from query-performance-prediction quantities such as softmax entropy, score gaps, and dispersion, and a gate of roughly one thousand parameters maps these signatures to per-query view weights. The fused ranking uses no relevance labels at inference, no re-ranking, and no fine-tuning of the retriever; the gate is trained leave-one-task-out. On the \textsc{Tempo} benchmark, the method improves all six retrievers we evaluate, from BERT encoders to 7B decoder retrievers, with the largest gains on the weaker backbones. The strongest retrievers reach $0.297$ and $0.303$ nDCG@10, and the per-query gain over the original query is significant under a paired bootstrap ($p<0.001$). A per-query oracle reaches $0.364$ against our realized $0.297$, exposing headroom that identifies per-query view selection as a concrete next step.
J. Holdcroft, Abdelrahman Abdallah, Adam Jatowt· 0 citations
We use cookies to run the site and, with your consent, for analytics and to show ads.
See our Cookie Policy.
