Query-Document Interaction Biases

• Query-document interaction biases are systematic distortions in retrieval systems caused by independent encoding and compression, leading to preferential treatment of early, repeated, or literal matching signals.
• They manifest across architectural, content, and contextual dimensions, impacting dense retrieval methods, pseudo-query generation, and user click behavior to reduce factual performance.
• Intervention strategies like dual cross-encoders and query-aware re-ranking demonstrate that mitigating bias requires explicit modeling of query-document interplay rather than simple corrections.

Query-document interaction biases are systematic distortions that arise from the way information retrieval systems and their users represent, compare, rank, and judge query-document pairs. Across recent work, the term covers architectural biases in dense retrieval, such as independent query and document encoding and single-vector compression; content-sensitive biases, such as head, brevity, literal-matching, and repetition effects; context-sensitive biases on search engine result pages (SERPs), where nearby results alter click and usefulness judgments; and query-side biases introduced by query generation, rewriting, or recommendation. The same literature also shows that evaluation pipelines inherit analogous distortions: LLM relevance judgments can fail in localized semantic neighborhoods, and conversational responses can appear satisfactory even when they are unsupported or incomplete (Li et al., 2022, Fayyaz et al., 6 Mar 2025, Chen et al., 2023, Sinha et al., 2024, Mohtadi et al., 5 Jan 2026).

1. Architectural origins of interaction bias

A central structural bias in vanilla dense retrieval is that the query and document are encoded independently and compared only after separate compression. In the standard dual-encoder pattern, retrieval is written as

$s(q,d)=DE_q(q)\cdot DE_d(d),$

which is efficient but forces the document tower to learn a single, query-agnostic embedding that must work for all query types. This creates a bias toward independent representations, weak or no query-document interaction, and a strong compression bottleneck for documents (Li et al., 2022).

Controlled analyses of dense retrievers show that this structural bias is not merely theoretical. In paired-document experiments built from Re-DocRED, retrievers favored shorter documents, early positions, repeated entities, and literal matches while ignoring whether the answer was present. When multiple biases were combined, models selected the answer-containing document in less than 10% of cases over a synthetic biased document without the answer, and in downstream RAG the retrieval-preferred poisoned document caused a 34% performance drop than providing no documents at all (Fayyaz et al., 6 Mar 2025). This establishes that dense retrieval can be dominated by heuristic interaction cues rather than factual evidence.

Position sensitivity provides a second major architectural regularity. The “Myopic Trap” study defined positional bias as a head bias in which retrieval quality drops when answer-bearing content is shifted from the beginning toward the middle or end of a document, even though the document’s semantics are unchanged. Embedding-based retrievers showed consistent degradation under these perturbations, and ColBERT-style late-interaction models also degraded, although bge-m3-colbert was more robust than bge-m3-dense under the same base model. By contrast, BM25 and rerankers remained largely unaffected, and representation analysis showed that the cosine similarity between full-document embeddings and beginning-segment embeddings was typically higher than the similarity to middle or end segments (Zeng et al., 20 May 2025).

These findings imply that query-document interaction bias is often encoded before any explicit relevance judgment is made. In single-vector systems, the document representation itself may already privilege early, short, repeated, or literally matching material. This suggests that many observed ranking errors are not isolated failures of scoring, but consequences of representation geometry and pooling.

2. Interaction-aware architectures and the reallocation of bias

One response to query-agnostic dense retrieval is to move part of the interaction into document encoding. The “dual cross encoder” replaces the standard document tower with a cross-encoder-like document encoder while keeping the query encoder as a standard dual encoder: $$q = DE_q(q), \qquad d = CE_d(q+d), \qquad s(q,d)= q\cdot d.$$ To preserve efficiency, the document is encoded offline with pseudo-queries generated from the document itself. For each document $d_j$, a T5-based query generation model samples 10 pseudo-queries with top-$k$ sampling, produces query-informed views $\mathbf{d}_j^i = CE_d(q_j^i + d_j)$, and scores at retrieval time with

$s(q,d_j) = \max_i \mathbf{q}^T\mathbf{d}_j^i.$

The method is therefore compatible with maximum inner product search while introducing deep query-document interaction into document encoding (Li et al., 2022).

This design does not eliminate bias; it reallocates it. The query encoder remains query-only, whereas the document encoder becomes query-conditioned, producing an explicit asymmetry between the two towers. Documents are represented through generated queries rather than from themselves alone, creating a pseudo-query inductive bias and a multi-view bias. The paper is also explicit about dependence on pseudo-query quality and a training-inference discrepancy: gold queries are available during training, but inference relies on generated pseudo-queries (Li et al., 2022).

Earlier interaction models made a related move by enriching the matching function itself. “Deep Relevance Ranking Using Enhanced Document-Query Interactions” retained DRMM’s query-term-wise scoring but injected context-sensitive BiLSTM encodings, PACRR-style convolutional n-gram features, attention-based document-aware query-term representations in ABEL-DRMM, pooled similarity statistics in POSIT-DRMM, and multi-view combinations of contextual, context-insensitive, and exact-match inputs. On BIOASQ, MAP increased from 49.3 for DRMM and 49.1 for PACRR to 50.7 for POSIT-DRMM and 51.0 for POSIT-DRMM+MV; on TREC ROBUST 2004, MAP increased from 25.6 for DRMM and 25.8 for PACRR to 27.0 for POSIT-DRMM and 27.2 for POSIT-DRMM+MV (McDonald et al., 2018). These results indicate that interaction design changes what counts as evidence: isolated lexical overlap, local n-gram alignment, context-sensitive meaning, or repeated support.

Recent listwise and multimodal systems extend the same principle. jina-reranker-v3 places the query and multiple documents in a single causal transformer context, extracts contextual document embeddings from a special token near the end of each document, and scores with cosine similarity. The model mitigates independent-encoding bias and enables cross-document influence, but the shared causal context introduces order sensitivity, positional effects, truncation dependence, and query-document asymmetry. In its ordering study on BEIR, random order scored 62.54, descending relevance order 61.94, and ascending order 61.52, which the paper interprets as reasonable robustness rather than order invariance (Wang et al., 29 Sep 2025). In visual document retrieval, Argus makes the document representation explicitly query-dependent, $\mathbf{D}(q)$, by routing page regions through a query-aware Mixture-of-Experts while retaining ColBERT-style MaxSim scoring. The same page is therefore not represented identically for a table lookup, a chart question, and a layout-sensitive evidence request (Abdallah et al., 3 Jun 2026).

A common misconception is that more interaction always means less bias. The literature is more precise: richer interaction usually mitigates query-agnostic isolation bias, but it can introduce new asymmetries, dependence on pseudo-queries or prompt order, and competition effects within the representation itself.

3. Contextual and cognitive biases on SERPs and in user judgments

Query-document interaction is also shaped by how results are presented to humans. A document-level decoy instance on a SERP is defined as a pair $(t,d)$ in which the target and decoy are content-similar but not identical, the target has higher quality, and the two results appear close together in rank: $$S_{\min} \leq \text{similarity}(t,d) \leq S_{\max}, \qquad \text{quality}(t) > \text{quality}(d), \qquad |rank(t)-rank(d)| \leq \Delta_{\mathrm{rank}}.$$ Using THUIR2016 and THU-KDD, the decoy-effect pilot study found that has_decoy increased click likelihood with coefficients 0.363 and 0.217, increased usefulness with coefficients 0.136 and 0.156, and increased browsing duration by about 1.9 seconds, although the dwell-time effect was not statistically significant. The strongest effects were therefore on click likelihood and usefulness ratings, even after controlling for rank, task, and user identity (Chen et al., 2023).

Whole-session search on debated topics exposes a broader interaction between user cognition and SERP composition. In simulated three-query sessions with 1,321 participants, most attitude change occurred in the first query. Ranking-biased SERPs produced the highest normalized confirming-click rate in the first query, with ClickProp 0.65 in the R condition versus 0.50 in Bal, while Counter-R and Obf increased ClickNum, ClickRank, and ClickDepth, consistent with users digging deeper for confirming information when top-ranked results were less aligned with their prior attitudes. The study also found that prior SERP bias altered later perceived familiarity and that low-openness users were more sensitive to the position of biased SERPs within the session (Wang et al., 2024).

Bias can also enter explicit relevance judgments. For six bias-sensitive queries, retrieved documents were manipulated to contain female or male gender indications while being relevant or non-relevant. In the gender-agnostic setting, the strongest trend for relevant documents occurred in Domestic Work, where the female-content relevant document was rated 2.10 versus 1.76 for the male-content version, with $p = 0.053$. For non-relevant documents, Appearance showed a statistically significant effect in the opposite direction: the male-content non-relevant document was rated 1.00 versus 0.70 for the female-content version, with $$q = DE_q(q), \qquad d = CE_d(q+d), \qquad s(q,d)= q\cdot d.$$0. In the gender-specific setting, participant gender did not significantly influence relevance judgments (Krieg et al., 2022).

Academic search engines show a related but system-side pattern. An audit of Google Scholar and Semantic Scholar using benefit-biased and risk-biased queries found that query bias positively predicted abstract valence, with $$q = DE_q(q), \qquad d = CE_d(q+d), \qquad s(q,d)= q\cdot d.$$1. The technology domain showed larger disparities than the health domain, with a significant interaction $$q = DE_q(q), \qquad d = CE_d(q+d), \qquad s(q,d)= q\cdot d.$$2, and Semantic Scholar produced smaller disparities overall than Google Scholar, with $$q = DE_q(q), \qquad d = CE_d(q+d), \qquad s(q,d)= q\cdot d.$$3 (Kacperski et al., 2023). This demonstrates that confirmation-biased query framing can be reinforced even in scholarly search.

Not all order-dependent interaction effects are large. In explanation-based interactive learning, within-session order effects had a limited, though significant, impact on agreement with the model, while between-session effects on agreement were not significant and subjective trust did not differ across conditions. This is an important contrast: some interaction biases materially alter clicks and usefulness on SERPs, whereas others remain small and inconsistent in debugging-oriented settings (Pesenti et al., 4 Dec 2025).

4. Query formulation, generation, and rewriting as sources of bias

Bias analysis depends critically on the query distribution used to probe a system. In retrievability studies, document accessibility is quantified as

$$q = DE_q(q), \qquad d = CE_d(q+d), \qquad s(q,d)= q\cdot d.$$4

and inequality over the collection is summarized with the Gini coefficient. However, comparisons across SQ1, SQ2, SQ3, RSQ, and AOL showed minimal or negligible correlation between retrievability scores derived from artificial queries and those derived from query logs. On Robust04, for example, AOL versus SQ2 yielded Pearson $$q = DE_q(q), \qquad d = CE_d(q+d), \qquad s(q,d)= q\cdot d.$$5 and Kendall $$q = DE_q(q), \qquad d = CE_d(q+d), \qquad s(q,d)= q\cdot d.$$6, while AOL versus RSQ yielded Pearson $$q = DE_q(q), \qquad d = CE_d(q+d), \qquad s(q,d)= q\cdot d.$$7 and Kendall $$q = DE_q(q), \qquad d = CE_d(q+d), \qquad s(q,d)= q\cdot d.$$8. Similar weak alignments appeared on WT10g and Wikipedia, and all differences were statistically significant with $$q = DE_q(q), \qquad d = CE_d(q+d), \qquad s(q,d)= q\cdot d.$$9. The paper therefore argues that query generation is not a neutral preprocessing step in retrievability or bias estimation (Sinha et al., 2024).

Query rewriting in RAG complicates this picture further. “Masking or Mitigating?” distinguishes query-document interaction biases, which can in principle be altered from the query side, from document encoding biases, which are structural preferences of the document encoder. On ColDeR, the baseline mean $d_j$0 across retriever-bias pairs was $d_j$1; simple Rewrite reduced this to 4.02, a 53.9% reduction, outperforming HyDE, HyDE-CPT, Q2D, and Q2D-CPT in aggregate. Yet the mechanistic analysis showed that Rewrite often achieved this reduction through increased score variance rather than genuine decorrelation from bias-inducing features, whereas pseudo-document methods could reduce sensitivity more directly. On the adversarial Foil subset, Rewrite barely improved outcomes, while HyDE-CPT raised CoCoDR from 2.0% to 26.0% (Goyal et al., 7 Apr 2026). The practical implication is that a lower aggregate bias statistic does not necessarily imply robust retrieval under compounded bias.

Balanced query recommendation treats the query itself as a lever for changing the demographic and geographic composition of retrieved documents. In the Wikipedia case study, the original query “Politics” retrieved results with very limited gender labeling, with only 1 out of 20 retrieved pages having a gender label. Candidate reformulations changed that distribution: “religion and culture influence politics” produced a geo entropy score of 1, indicating a completely dissimilar geographic distribution from the original query’s results, while “liberal politicians debate” yielded a gender entropy score of 0.248 and surfaced both male and female political biographies (Mishra et al., 28 Aug 2025). This shows that subtle linguistic variation can alter the protected-attribute distribution of the retrieved set even when the topical theme is broadly stable.

A more strategic formulation appears in “Querying with Conflicts of Interest.” There, a user’s true intent $d_j$2 is expressed through a submitted query $d_j$3, while the source may reinterpret it as $d_j$4 to serve its own incentives. The framework models a user strategy $d_j$5, a source strategy $d_j$6, and a Bayesian posterior

$d_j$7

The paper then gives algorithms for detecting whether influential interaction is possible, identifying untrustworthy tuples, and reformulating queries with rank constraints to recover more relevant results from biased sources (Aryal et al., 5 Mar 2026). This formalizes a limit case in which query-document interaction bias is intentional rather than incidental.

5. Diagnosing localized failures in evaluation and response generation

The same need for joint query-document analysis appears in evaluation. “Query-Document Dense Vectors for LLM Relevance Judgment Bias Analysis” argues that relevance is a property of the query-document pair and embeds the pair as a single dense vector

$d_j$8

using INSTRUCTOR-XL and the instruction “Judge the document’s relevance to the query for ad-hoc retrieval.” Q-D embeddings are clustered with HDBSCAN, and agreement between LLM and human labels is measured with Gwet’s AC1 rather than Cohen’s $d_j$9, because $k$0 is unstable under label imbalance. The paper’s key diagnostic is Cluster-based Agreement Variation,

$k$1

which localizes query-specific instability across semantic neighborhoods. On TREC Deep Learning 2019 and 2020, disagreement was concentrated in specific clusters rather than distributed randomly, with recurring failures in definition-seeking, policy-related, ambiguous, and domain-sensitive queries. The dominant error patterns were under-recall from over-strict filtering and over-inclusion from superficial lexical matching (Mohtadi et al., 5 Jan 2026).

Conversational information-seeking exhibits a related mismatch between actual support and perceived quality. In two crowdsourcing studies over TREC CAsT queries, users evaluated responses under controlled manipulations of factual correctness, source validity, diversity, and balance. In the answerability study, controlled factual correctness and source presence or validity had no statistically significant effect on user ratings, indicating that users did not reliably detect factual errors or unsupported claims. In the viewpoints study, users were better at detecting diversity and balance problems. Satisfaction correlated with Factual Correctness at 0.634 and Confidence in Answer Accuracy at 0.660, but it correlated even more strongly with Diversity at 0.720, Transparency at 0.727, and Balance at 0.785 (Łajewska et al., 2024). Fluent, diverse-seeming responses can therefore mask factual weakness.

A different decomposition appears in query-focused multi-document summarization. Instead of treating query-document interaction as a single relevance score, the framework separates query relevance, evidence or answer likelihood, and centrality. A simple retrieval stage produces top candidates; a BERT-based evidence estimator, trained with distant supervision from WikiQA, TrecQA, and SQuAD 2.0, estimates whether a sentence or passage actually answers the query; and a query-aware LexRank variant combines centrality with the evidence distribution. The paper reports that QuerySum$k$2 achieved the best ROUGE-1, QuerySum$k$3 the best ROUGE-2 and ROUGE-SU4, and ablations showed that removing evidence or centrality hurt performance (Xu et al., 2020). This operationalizes a broader principle: topical similarity, answerability, and summary worthiness are distinct interaction phenomena and should not be collapsed into one signal.

6. Conceptual synthesis and implications

The literature does not use “bias” in a single sense. One strand studies representational bias: independent encoding, single-vector compression, head bias, brevity bias, literal matching, and repetition preferences in retrievers (Fayyaz et al., 6 Mar 2025, Zeng et al., 20 May 2025). A second studies comparative and cognitive bias in human interaction, including decoy effects, confirmation bias, stereotype-laden relevance judgments, and order effects (Chen et al., 2023, Wang et al., 2024, Krieg et al., 2022). A third studies query-side and evaluative bias, where artificial query generation distorts retrievability estimates, rewriting changes measured bias without necessarily improving robustness, and LLM judges or conversational responses fail in systematic but localized ways (Sinha et al., 2024, Goyal et al., 7 Apr 2026, Mohtadi et al., 5 Jan 2026, Łajewska et al., 2024).

Several misconceptions follow from conflating these levels. Query-document interaction bias is not reducible to social unfairness alone; it also includes architectural regularities such as head bias and literal-match preference. Nor is it equivalent to rank bias: the decoy-effect study showed that position-bias correction is not enough, because local comparison still shifts clicks and usefulness (Chen et al., 2023). Likewise, stronger interaction does not simply “remove” bias. Query-conditioned encoders, shared-context rerankers, and region-aware late interaction reduce query-agnostic isolation bias, but they introduce asymmetry, pseudo-query dependence, order sensitivity, or positional privilege inside the model (Li et al., 2022, Wang et al., 29 Sep 2025, Abdallah et al., 3 Jun 2026).

For evaluation and deployment, the cumulative implication is that average metrics often conceal where systems fail. Dense retrievers may surface superficially compatible but factually inferior documents; query logs and artificial queries can yield different retrievability landscapes; LLM judges can agree globally while failing badly in specific semantic clusters; and users may reward balance or fluency even when support is absent (Fayyaz et al., 6 Mar 2025, Sinha et al., 2024, Mohtadi et al., 5 Jan 2026, Łajewska et al., 2024). This suggests that robust analysis of query-document interaction requires joint representations, localized diagnostics, and explicit separation of relevance from adjacent constructs such as answerability, centrality, usefulness, and trust.

The broad trajectory of the field is therefore not the elimination of interaction bias, but its explicit modeling. Query-conditioned document views, multi-vector late interaction, cross-document contextualization, behavior-aware SERP analysis, bias-aware query generation, and cluster-level disagreement analysis all treat the query-document pair as a relational object rather than as two independently meaningful items. That shift does not make retrieval or evaluation bias-free. It does, however, make the sources, failure modes, and trade-offs of bias more visible, more measurable, and more open to targeted intervention.