Interaction Queries Overview

• Interaction queries are queries whose interpretation and execution are dynamically refined through user-system interaction to resolve ambiguity.
• They leverage formal models like uncertainty reduction and game theory to support iterative clarification and schema exploration in various applications.
• Empirical findings show that even a few user interactions significantly improve query accuracy and usability in complex data environments.

Interaction queries are queries whose interpretation, reformulation, execution, or answer timing is coupled to an interaction loop rather than fixed by a single submission. Across database systems, semantic question answering, information retrieval, tabular analysis, and multimodal agents, they arise when users do not precisely know the database structure or content, when natural-language questions are ambiguous or context-dependent, when ad-hoc query formulation over large schemas is difficult, or when the evidence needed for an answer appears asynchronously with respect to the query itself (McCamish et al., 2016, Zafar et al., 2020, Erbacher et al., 2022, Zhang et al., 23 Jun 2025). Recent work therefore treats interaction not as an interface afterthought but as part of the query semantics itself, including explicit frameworks for cooperative ambiguity resolution and for deciding when clarification, reformulation, or deferral is required (Gomm et al., 6 Nov 2025).

1. Scope and taxonomy

A central distinction in recent work is between query execution and query interpretation. In tabular data analysis, an actionable query interpretation is defined as the unique combination of a specific analytical procedure and the exact data scope, including entities, temporal bounds, and domain constraints (Gomm et al., 6 Nov 2025). On that basis, a cooperative query is one that provides sufficient specification, either explicitly or through reasonable inference, to identify at least one valid, actionable query interpretation, whereas an uncooperative query is underspecified to the degree that ambiguity is irresolvable (Gomm et al., 6 Nov 2025). This framing recasts ambiguity as part of cooperative interaction rather than as a purely technical defect.

Database interaction work formulates the same problem in behavioral terms. Users often do not precisely know the structure and content of databases, so their queries do not exactly reflect their information needs, and both users and DBMSs adapt during interaction rather than operating with fixed strategies (McCamish et al., 2016). In information retrieval, the corresponding issue is that initial search queries are often unclear or lack context, so clarification questions or clarification panels become part of the retrieval process rather than a separate usability layer (Erbacher et al., 2022). In semantic question answering, uncertainty appears as a space of candidate semantic interpretations over a knowledge graph, especially for complex questions involving multiple entities and relations (Zafar et al., 2020).

Interaction queries also include temporally misaligned settings. AViLA defines Query-Evidence Asynchrony as the mismatch between when a user issues a query and when supporting evidence appears in a streaming video, distinguishing evidence-before-query, evidence-during-query, and evidence-after-query cases (Zhang et al., 23 Jun 2025). A related context-sensitive form appears in semantic parsing for interactions, where each utterance must be mapped to an executable formal query while referencing previous utterances and previously generated queries (Suhr et al., 2018).

2. Formal models of interactive querying

One major line of work treats interaction as uncertainty reduction over a hypothesis space. IQA introduces the question interpretation space $QIS$, whose elements are complete query interpretations, and ranks interaction options with the Option Gain metric

$OG(IO) = usability(IO)^{\omega} \times IG(IO),$

where $usability(IO)$ is in $[0,1]$, $IG(IO)$ is the expected entropy reduction, and $\omega$ controls the weight of usability (Zafar et al., 2020). The interaction process is explicitly described as resembling a cost-sensitive decision tree, with user feedback pruning candidate semantic queries (Zafar et al., 2020).

A second line models interaction as a game between adaptive agents. In the signaling-game framework for DBMS querying, intents $\{e_1,\ldots,e_m\}$, queries $\{q_1,\ldots,q_n\}$, and interpretations or results $\{s_1,\ldots,s_o\}$ are linked by a user strategy $U$ and a DBMS strategy $OG(IO) = usability(IO)^{\omega} \times IG(IO),$0, both represented as row-stochastic matrices (McCamish et al., 2016). The shared objective is the expected payoff

$OG(IO) = usability(IO)^{\omega} \times IG(IO),$1

and both sides are updated with Roth-Erev reinforcement learning (McCamish et al., 2016). The paper reports that the expected payoff $OG(IO) = usability(IO)^{\omega} \times IG(IO),$2 is a submartingale and converges almost surely, formalizing the emergence of a common language for intents and queries (McCamish et al., 2016).

Context-dependent semantic parsing adds an explicit interaction state. The ATIS model maintains a discourse state $OG(IO) = usability(IO)^{\omega} \times IG(IO),$3 that is updated after each turn, conditions the utterance encoder on $OG(IO) = usability(IO)^{\omega} \times IG(IO),$4, and augments generation with copying of sub-sequences from previously predicted queries (Suhr et al., 2018). The result is a combination of implicit reference modeling through attention over prior utterances and explicit reference modeling through segment copying, which directly targets ellipsis and cross-turn reuse (Suhr et al., 2018).

An older but conceptually related formalization appears in interactive IR based on quantum probabilistic representations. There, the first interaction of retrieval—the user typing a query—is modeled as a query density operator $OG(IO) = usability(IO)^{\omega} \times IG(IO),$5, documents are represented as subspaces with projector $OG(IO) = usability(IO)^{\omega} \times IG(IO),$6, and relevance is $OG(IO) = usability(IO)^{\omega} \times IG(IO),$7 (Piwowarski et al., 2010). That work explicitly argues that representing queries and documents as multidimensional objects is a step toward “truly” interactive IR systems (Piwowarski et al., 2010).

3. Interactive clarification and refinement mechanisms

In semantic question answering, IQA inserts interaction directly into the SQA pipeline. The architecture comprises a Shallow Parser, Entity Linker, Relation Linker, and Query Builder, each of which may use multiple sub-tools to increase candidate recall (Zafar et al., 2020). The user is shown interaction options drawn from four categories—nugget interpretations, superclasses or types, answer types, and complete queries—and can confirm or reject them while the system recalculates Option Gain after each update (Zafar et al., 2020). Interaction terminates when the user accepts a complete query or reaches a preset interaction limit (Zafar et al., 2020).

In information retrieval, fully simulated multi-turn clarification has been used to study dynamic query refinement. The framework of “Interactive Query Clarification and Refinement via User Simulation” starts from an initial query $OG(IO) = usability(IO)^{\omega} \times IG(IO),$8, generates a diverse set of candidate reformulations with T5 and diverse beam search, and then repeatedly presents $OG(IO) = usability(IO)^{\omega} \times IG(IO),$9 candidate queries, often $usability(IO)$0, to a user agent that selects the one with highest similarity to the intent embedding (Erbacher et al., 2022). A conditional ranking model with a Hierarchical RNN uses the sequence of positive and negative feedback pairs $usability(IO)$1 to update the ranking over subsequent turns (Erbacher et al., 2022). This makes the clarification process history-aware and dynamic rather than log-replayed (Erbacher et al., 2022).

InteracSPARQL applies the same interaction principle to semantic web querying. It first parses SPARQL into an AST, produces rule-based structured natural language explanations from the AST, refines these explanations with an LLM, and then enters an interactive refinement loop in which execution results, explanations, and targeted feedback are jointly used to update the query (Jian et al., 3 Nov 2025). Entity and property errors are resolved through knowledge-graph tool calls, and only the components flagged as problematic are revised, preserving validated parts of the query (Jian et al., 3 Nov 2025). The design goal is not merely NL-to-SPARQL generation, but iterative query refinement through explanation.

A related LLM-mediated mechanism appears in medical visualization. “Changing the Paradigm from Dynamic Queries to LLM-generated SQL Queries with Human Intervention” replaces traditional sliders, check boxes, and drop-downs with natural-language requests that are translated into editable SQL, paired with a natural-language explanation and assumptions, and then executed to drive the visualization (Assor et al., 11 Sep 2025). The paper also reintroduces dynamic queries on demand by extracting SQL subtrees into widgets for rapid, incremental, reversible control over selected filters (Assor et al., 11 Sep 2025). This preserves human intervention while moving query formulation into an editable symbolic form.

4. Query formulation over large conceptual schemas

A distinct tradition studies interaction queries in conceptual-schema environments where the central difficulty is not ambiguity in language but the scale and structure of the schema. Query by Navigation decomposes formulation into an explorative phase and a constructive phase (Proper, 2021). In the explorative phase, users build linear path expressions by navigating the schema through refinements and enlargements; in the constructive phase, they combine the resulting paths using concatenation, union, intersection, and selection (Proper, 2021). The underlying query-by-navigation graph is formalized as $usability(IO)$2, and navigation is treated as traversal over partial queries rather than over instance data alone (Proper, 2021).

Point to Point Queries address the case in which users know which object types matter but not how they are related. The schema is modeled as a labeled undirected graph $usability(IO)$3, candidate paths are ranked by a badness function that balances conceptual importance and path length, and the system returns the “best” paths first with a MORE option for incremental disclosure (Proper, 2021). Because enumerating all paths is intractable, the paper uses incremental best-first search with schema reduction and clustering to prune irrelevant portions of the graph before path discovery (Proper, 2021).

Spider queries provide a complementary mechanism for broad contextual expansion around a known object type. Starting from an object type, the system recursively expands to all directly related object types, producing a rooted tree that “fans out” over the conceptual schema (Proper, 2021). Users can then prune branches that are not relevant, and any branch can itself become the root of a new spider query (Proper, 2021). The construction avoids cycles by ensuring that no object type appears twice on the same root-to-node path, and the resulting structure can be translated into path expressions suitable for SQL or natural-language verbalization (Proper, 2021).

Taken together, these mechanisms shift query formulation from direct symbolic authoring toward controlled schema exploration. They are especially suited to ad-hoc querying in large conceptual schemas, where users may know domain concepts but not schema topology (Proper, 2021, Proper, 2021, Proper, 2021).

5. Empirical findings on effectiveness, usability, and real user behavior

Empirical work consistently reports that a small number of interactions can materially improve query quality. In IQA, user studies on LC-QuAD show that IQA-OG achieves up to 20 percentage points improvement in $usability(IO)$4 score over the best non-interactive baseline, that IQA-OG gives a 10 point higher $usability(IO)$5 score than IQA-IG, and that both IQA-OG and IQA-IG require only about $usability(IO)$6–$usability(IO)$7 interactions on average (Zafar et al., 2020). Users rated IQA-OG at $usability(IO)$8 versus $usability(IO)$9 for IQA-IG, with the difference reported as statistically significant at $[0,1]$0 (Zafar et al., 2020).

The fully simulated IR framework reports the same qualitative trend. On the MS MARCO 2020 passage test set, “Interact. + KMeans” rises from $[0,1]$1 at $[0,1]$2 turns to $[0,1]$3 at $[0,1]$4 turn and $[0,1]$5 at $[0,1]$6 turns in MRR@10, while “Interact. + Random Sampl.” rises from $[0,1]$7 to $[0,1]$8 and then $[0,1]$9 (Erbacher et al., 2022). The reported upper bound remains the oracle “Best Reformulation” at $IG(IO)$0, so interaction is framed as narrowing rather than eliminating the gap between naive ranking and best possible reformulation (Erbacher et al., 2022).

InteractComp shows, however, that current search agents still underuse interactive clarification. Across 210 expert-curated ambiguous queries in 9 domains, evaluated over 17 models, the best model reaches only $IG(IO)$1 accuracy, while OpenAI o3 reaches $IG(IO)$2 when given complete context (Deng et al., 28 Oct 2025). Forced interaction changes the picture: for GPT-5, increasing interactions from $IG(IO)$3 to $IG(IO)$4 raises accuracy from $IG(IO)$5 to $IG(IO)$6, and the paper identifies systematic overconfidence rather than reasoning deficits as the main bottleneck (Deng et al., 28 Oct 2025). It also reports that interaction capability remained stagnant at $IG(IO)$7–$IG(IO)$8 over 15 months while search performance on complete queries improved seven-fold (Deng et al., 28 Oct 2025).

User-query studies indicate that real interaction behavior is heterogeneous and often weakly specified. In the legal-aid dataset of 3,847 queries from 1,252 users, 29.95% of queries contained factual information, 64.93% sought legal information rather than advice, and 28.57% shaped or controlled the model’s answer (Kuk et al., 3 Jan 2025). Only 129 queries, or 3.35%, fit the “personalized actionable advice” extreme, and only 117 queries, or 3.04%, fit the “pure legal look-up” extreme; 93.61% were hybrid or intermediate (Kuk et al., 3 Jan 2025). This empirical pattern is consistent with the claim that query interfaces must accommodate mixed intentions rather than assume either pure lookup or fully specified consultation.

Evaluation methodology itself has become a research issue. The tabular-analysis study examines 15 popular datasets and reports an uncontrolled mixing of unambiguous, cooperative, and uncooperative queries, which confounds execution accuracy with interpretation capability and undermines single-gold-answer evaluation when multiple reasonable interpretations exist (Gomm et al., 6 Nov 2025). A plausible implication is that future evaluation of interaction queries must stratify by grounding requirements rather than collapse all query types into a single benchmark score.

6. Emerging directions and terminological extensions

Streaming multimodal agents extend interaction queries into time-critical settings. AViLA maintains query-agnostic memory, identifies evidence relevant to each ad-hoc query, and uses an evidence-grounded trigger to decide whether to respond or defer (Zhang et al., 23 Jun 2025). On AnytimeVQA-1K, AViLA with adversarial verification trigger reaches $IG(IO)$9 accuracy with a mean temporal offset of $\omega$0 seconds, compared with $\omega$1 and $\omega$2 seconds for Flash-VStream and $\omega$3 and $\omega$4 seconds for VideoLLM-Online (Zhang et al., 23 Jun 2025). This makes response timing part of query correctness.

Another direction treats queries as specifications from which interfaces themselves can be synthesized. PI2 takes an example sequence of analysis queries, parses them into Difftrees that encode systematic AST variations, maps these variations to interactions and visualizations, and generates full interfaces in $\omega$5–$\omega$6 seconds with a median of $\omega$7 seconds while scaling linearly with the number of queries (Chen et al., 2021). Here the interaction query is not only an object to be answered but also a design primitive for producing the interface that will answer future related queries (Chen et al., 2021).

The phrase “query interaction” has also acquired a different technical meaning in model architecture work. In camera-radar fusion, ConFusion introduces heterogeneous query interaction among image queries, radar queries, and world queries through QMix and QSwap, reaching $\omega$8 mAP and $\omega$9 NDS on the nuScenes validation set and $\{e_1,\ldots,e_m\}$0 mAP and $\{e_1,\ldots,e_m\}$1 NDS on the test set (Wu et al., 28 Apr 2026). In visual relationship detection, PST uses composite queries to model subject, predicate, object, and triplet-as-sum jointly, achieving $\{e_1,\ldots,e_m\}$2 R@100 on VRD and $\{e_1,\ldots,e_m\}$3 mAP on HICO-DET among single-stage models (Dong et al., 2021). This suggests that, in recent arXiv usage, the term has become polysemous: it may denote human-system interaction around query meaning, or interaction among learned query objects inside a neural architecture.

Across these strands, several design implications recur. Systems should distinguish when to automate and when clarification is required, expose system inferences so that users can overrule them, and support iterative refinement rather than single-shot execution (Gomm et al., 6 Nov 2025, Assor et al., 11 Sep 2025). At the interface layer, high-level interaction models such as Interacto, which reify user interactions as first-class entities and connect them to commands with built-in undo/redo support, indicate that robust handling of interaction queries depends not only on ranking or parsing models but also on the software abstractions through which interactions are assembled, tested, and reused (Blouin et al., 2021).