QUARK: Robust Retrieval under Non-Faithful Queries via Query-Anchored Aggregation

TL;DR

QUARK addresses retrieval under non-faithful queries by explicitly modeling uncertainty through multiple recovery hypotheses and aggregating their signals with a query-anchored scheme. It operates training-free on top of any base retriever, using a simple interpolation with a latent intent-aware max aggregator: $S_{agg}(q,d)=\alpha S(q,d)+(1-\alpha)\max_{h\in\mathcal{H}(q)}S(h,d)$. Across controlled simulations and BEIR benchmarks, QUARK consistently improves Recall, MRR, and nDCG without retraining, with strongest gains at moderate noise levels and when using an appropriate $\alpha$ in a high-range (e.g., $[0.6,0.9]$). The approach demonstrates that incorporating multiple intent-preserving interpretations and anchoring them to the original query can mitigate semantic drift and hypothesis hijacking, providing robust retrieval under challenging noisy queries with broad practical impact for real-world search systems.

Abstract

User queries in real-world retrieval are often non-faithful (noisy, incomplete, or distorted), causing retrievers to fail when key semantics are missing. We formalize this as retrieval under recall noise, where the observed query is drawn from a noisy recall process of a latent target item. To address this, we propose QUARK, a simple yet effective training-free framework for robust retrieval under non-faithful queries. QUARK explicitly models query uncertainty through recovery hypotheses, i.e., multiple plausible interpretations of the latent intent given the observed query, and introduces query-anchored aggregation to combine their signals robustly. The original query serves as a semantic anchor, while recovery hypotheses provide controlled auxiliary evidence, preventing semantic drift and hypothesis hijacking. This design enables QUARK to improve recall and ranking quality without sacrificing robustness, even when some hypotheses are noisy or uninformative. Across controlled simulations and BEIR benchmarks (FIQA, SciFact, NFCorpus) with both sparse and dense retrievers, QUARK improves Recall, MRR, and nDCG over the base retriever. Ablations show QUARK is robust to the number of recovery hypotheses and that anchored aggregation outperforms unanchored max/mean/median pooling. These results demonstrate that modeling query uncertainty through recovery hypotheses, coupled with principled anchored aggregation, is essential for robust retrieval under non-faithful queries.

Figures (5)

• Figure 1: The overview of QUARK.
• Figure 2: Prompt templates used with Qwen3-maxyang2025qwen3 to generate observed lyric-recall queries under three levels of non-faithfulness in the lyrics-retrieval simulation.
• Figure 3: Prompt for lyrics-retrieval hypothesis generation.
• Figure 4: Sensitivity of QUARK to $\alpha$ on the lyrics-retrieval simulation (L2), reported as changes relative to the baseline retriever ($\alpha=1.0$).
• Figure 5: Sensitivity of QUARK to $\alpha$ on BEIR benchmarks with BM25, reported as changes relative to the baseline retriever ($\alpha=1.0$).