Adaptive Query Refinement (AdaptQA)

• Adaptive Query Refinement (AdaptQA) is a dynamic strategy that iteratively refines queries based on evidence gaps, conversational shifts, and safety signals.
• It leverages control signals like topic assignments, complexity predictions, and reward models to trigger targeted sub-query generation and rewriting.
• AdaptQA is applied in diverse areas such as multi-hop QA, search, database repair, and safety systems to enhance response accuracy and reduce failure rates.

Adaptive Query Refinement (AdaptQA) denotes a class of mechanisms that modify a query, its sub-queries, its retrieval scope, or its execution plan as evidence, conversation state, or safety analysis evolves. In the cited literature, the term covers several closely related behaviors: evidence-driven sub-query generation in retrieval-augmented generation, topic-guided query rewriting in multi-hop QA, strategy routing based on predicted query complexity, conversational replanning under goal shifts, safety-preserving prompt rewriting, and test-time adaptation of embeddings or reformulations for downstream retrieval (asl et al., 25 Oct 2025, Rezaei et al., 2024, Jeong et al., 2024, Lee et al., 27 Sep 2025, Shen et al., 27 Aug 2025).

1. Definitions and scope

In retrieval-augmented generation, AdaptQA is defined as a dynamic alternative to a static “retrieve-then-read” pipeline. FAIR-RAG uses AdaptQA as the agentic component that receives explicit informational gaps from Structured Evidence Assessment (SEA) and turns them into targeted sub-queries, repeating the loop of retrieval, filtering, auditing, and refinement until SEA certifies sufficiency (asl et al., 25 Oct 2025). AT-RAG defines adaptive query refinement as dynamically tailoring both the retrieval scope and the query itself as evidence accumulates, operationalized through BERTopic-based topic assignment, chain-of-thought reasoning, answer grading, and query rewriting (Rezaei et al., 2024).

Outside classical RAG, the scope broadens. In EHR-ChatQA, AdaptQA is an interaction flow in which the user changes the search goal mid-conversation, forcing the agent to re-plan, resolve value mismatches, and synthesize a new SQL plan rather than merely adding constraints to an unchanged goal (Lee et al., 27 Sep 2025). In IntentionReasoner, AdaptQA is a safeguard principle: a guard model reasons about intent, assigns one of four safety labels, and selectively rewrites borderline prompts so that a downstream model can respond safely without excessive refusal (Shen et al., 27 Aug 2025). ReQueR uses query refinement as inference-time alignment, training a Refiner policy to rewrite ambiguous human questions into explicit logical decompositions that better elicit latent reasoning in frozen solvers (Zhou et al., 28 Apr 2026).

The literature therefore uses AdaptQA to denote a family of adaptive control mechanisms rather than a single algorithm. What remains common is that refinement is triggered by some diagnostic signal—evidence gaps, topic assignments, answer graders, complexity labels, reward-model scores, execution failures, user feedback, or safety labels—and that the refined query is judged by downstream utility rather than by surface similarity alone.

2. Common formal structure and control signals

A recurring formal pattern is a loop that alternates between query transformation and external validation. FAIR-RAG provides the clearest canonical formulation. Given a user query $Q$, SEA deconstructs it into required findings $F = D(Q) = \{f_1, f_2, \dots, f_M\}$. At iteration $t$, curated evidence accumulates as $E(t) = E(t-1) \cup Filter(D_t, Q)$ with $E(0)=\varnothing$. SEA then computes the confirmed findings $C(t)$ and the explicit gap set $G(t)=F \setminus C(t)$, and returns sufficiency exactly when $C(t)=F$; AdaptQA then maps $(E(t), Q, G(t), C(t))$ to targeted sub-queries $S(t)$ that close the remaining gaps (asl et al., 25 Oct 2025).

Other systems instantiate the same controller pattern with different diagnostics. AT-RAG uses topic assignment and an Answer Grader composed of UsefulnessGrader and HallucinationGrader; if the current answer is not useful or is hallucinated, the controller applies $F = D(Q) = \{f_1, f_2, \dots, f_M\}$0 and continues (Rezaei et al., 2024). Adaptive-RAG moves the control decision one level up: a T5-Large classifier predicts a complexity label $F = D(Q) = \{f_1, f_2, \dots, f_M\}$1 and routes the query to no retrieval, single-step retrieval, or iterative multi-step retrieval, thereby deciding whether query refinement is needed at all (Jeong et al., 2024). AdaRewriter performs Best-of-$F = D(Q) = \{f_1, f_2, \dots, f_M\}$2 conversational reformulation and selects the final standalone query by $F = D(Q) = \{f_1, f_2, \dots, f_M\}$3 with $F = D(Q) = \{f_1, f_2, \dots, f_M\}$4, where $F = D(Q) = \{f_1, f_2, \dots, f_M\}$5 is a reward model trained from retrieval outcomes (2506.01381). AdaQR similarly aligns a rewriter without passage labels by assigning each rewrite candidate a reward $F = D(Q) = \{f_1, f_2, \dots, f_M\}$6, where the retriever weights top-$F = D(Q) = \{f_1, f_2, \dots, f_M\}$7 passages and a fixed scorer computes answer log-probability conditioned on the original conversational turn and each passage (Zhang et al., 2024).

These formulations differ in granularity, but their control signals are structurally analogous. FAIR-RAG uses explicit gap sets; AT-RAG uses topic filters and grader outputs; Adaptive-RAG uses predicted complexity; AdaRewriter and AdaQR use learned or weakly supervised reward signals. The literature suggests that AdaptQA is best understood as a closed-loop controller over query state, not merely as a rewrite operator.

3. Domain-specific embodiments

AdaptQA has been instantiated across at least five distinct technical settings. In multi-hop and ambiguous QA, it appears as sub-query synthesis, retrieval diversification, or strategy routing. AQA formulates adaptive QA as a contextual multi-armed bandit in which the context is a question feature vector and each arm is a communication graph over agents such as NoR, OneR, and IRCoT; the framework learns which answering strategy to invoke for each query type (Hoveyda et al., 2024). DIVA, designed for ambiguous questions, generates pseudo-interpretations, retrieves per interpretation, verifies whether the final passage pool is Useful, PartialUseful, or Useless, and then adapts the answering branch accordingly, including a closed-book fallback when retrieval is judged Useless (In et al., 2024).

In search and query expansion, AdaptQA often takes the form of test-time selection or domain-adaptive expansion. Automatic in-domain exemplar construction with multi-LLM refinement builds domain-specific demonstration pools via BM25→MonoT5 pseudo-relevance feedback, selects diverse exemplars by clustering, generates two expansions with heterogeneous LLMs, and refines them into a single expansion before re-retrieval (Li et al., 9 Feb 2026). Task-Adaptive Embedding Refinement does not rewrite text directly; instead, it optimizes the query embedding $F = D(Q) = \{f_1, f_2, \dots, f_M\}$8 by minimizing $F = D(Q) = \{f_1, f_2, \dots, f_M\}$9 over a small candidate set, using yes/no judgments from a teacher LLM as the target distribution (Gera et al., 12 May 2026). The interactive Query Generation Assistant provides a human-in-the-loop variant for multilingual search: users edit prompts, mark relevant results, and inject retrieved examples back into the prompt so that later generations remain aligned with the intended facet (Dhole et al., 2023).

In database and formal query settings, AdaptQA becomes execution-aware repair. SEA-SQL separates initial SQL generation, Adaptive Bias Elimination, and Dynamic Execution Adjustment: the first uses a semantic-enhanced schema, the second uses a fine-tuned Mistral-7B to remove model-specific bias from pre-SQL conditioned on execution results, and the third iteratively executes, reflects, and corrects until the SQL becomes executable (Li et al., 2024). InteracSPARQL similarly parses SPARQL into an AST, derives rule-based natural-language explanations, refines those explanations with an LLM, and then supports iterative repair through user feedback or LLM self-refinement with entity/property lookup tools (Jian et al., 3 Nov 2025). In EHR-ChatQA, AdaptQA is the harder of two conversational evaluation flows because the target goal itself may change, requiring the agent to abandon stale constraints and synthesize a new SQL plan consistent with the updated clinical objective (Lee et al., 27 Sep 2025).

A further branch concerns reasoning elicitation and safety. ReQueR trains a Refiner policy by reinforcement learning to output explicit logical decompositions in a constrained format, using frozen solvers as the environment and an Adaptive Solver Hierarchy as curriculum (Zhou et al., 28 Apr 2026). IntentionReasoner uses intent reasoning and selective rewriting to neutralize harmful intent in borderline prompts while preserving benign sub-goals, rather than treating safety as a pure reject/allow decision (Shen et al., 27 Aug 2025).

4. Empirical record

Reported results show that AdaptQA-style control is valuable when the query distribution is heterogeneous, but the magnitude and character of the gains depend strongly on domain, task, and failure mode. Representative figures from FAIR-RAG, AT-RAG, Adaptive-RAG, AdaRewriter, EHR-ChatQA, Task-Adaptive Embedding Refinement, and IntentionReasoner are summarized below (asl et al., 25 Oct 2025, Rezaei et al., 2024, Jeong et al., 2024, 2506.01381, Lee et al., 27 Sep 2025, Gera et al., 12 May 2026, Shen et al., 27 Aug 2025).

Framework	Benchmark or setting	Reported result
FAIR-RAG 3	HotpotQA / 2WikiMultiHopQA / MusiQue	F1 = 0.453 / 0.320 / 0.264; HotpotQA is +8.3 absolute over the strongest iterative baseline
AT-RAG (GPT-4o)	2WikiMultiHopQA / HotpotQA / MuSiQue	Overall = 6.57 / 7.61 / 4.52
Adaptive-RAG (GPT-3.5)	Average across six QA datasets	F1 = 50.91, Steps = 1.03, Time = 1.46
AdaRewriter (Llama3.1-8B, $t$0)	TopiOCQA	Sparse: MRR 30.7, NDCG@3 28.8, R@10 51.3; Dense: MRR 40.3, NDCG@3 39.7, R@10 61.9
EHR-ChatQA AdaptQA (o4-mini)	Combined AdaptQA tasks	SR-5 = 43.8%, Pass@5 = 78.1%, Pass$t$1 = 15.6%, Gap-5 = 62.5%
Task-Adaptive Embedding Refinement	Qwen3-Embedding-0.6B average across tasks	MAP .55 → .66 (+19.2%)
IntentionReasoner IR-7B	Harmfulness detection suites	Average F1 ≈ 99.4, ASR ≈ 1.2%, ORR ≈ 0.0%

Several cross-cutting patterns recur. First, multi-hop QA benefits from adaptive iteration, but only up to a point. FAIR-RAG reports that best average answer rank typically occurs at $t$2–$t$3 on multi-hop tasks, whereas $t$4 degrades quality despite added tokens and calls; on HotpotQA, average per-query cost rises from 9,787 tokens at $t$5 to 17,281 at $t$6 and 20,299 at $t$7 (asl et al., 25 Oct 2025). Adaptive-RAG shows the same budget-quality principle at the routing level: for GPT-3.5, Adaptive-RAG achieves average F1 50.91 with Steps 1.03 and Time 1.46, while always using the multi-step branch gives F1 50.87 with Steps 2.81 and Time 3.33 (Jeong et al., 2024).

Second, robustness is a separate axis from optimistic success. EHR-ChatQA reports high Pass@5 but much lower Pass$t$8 on AdaptQA, and the gap is explicitly framed as a robustness gap; AdaptQA requires on average 5.5 user interactions and 10.2 tool interactions per task, compared with 4.9 and 7.0 for IncreQA (Lee et al., 27 Sep 2025). Third, gains are not limited to symbolic or retrieval-heavy settings. Task-Adaptive Embedding Refinement reports relative MAP improvements of up to +25.5% on CLINC150 and +24.5% on RealScholar for Qwen3-Embedding-0.6B, indicating that refinement can act directly on the embedding representation rather than on query text (Gera et al., 12 May 2026). In safeguard settings, IntentionReasoner shows that selective query rewriting can simultaneously reduce over-refusal and maintain strong adversarial robustness, with IR-7B reporting near-zero ORR and low jailbreak ASR (Shen et al., 27 Aug 2025).

5. Failure modes, trade-offs, and common misconceptions

The dominant failure modes differ by architecture, but they are rarely random. FAIR-RAG’s 200-case error analysis attributes 63.5% of failures to component-level issues—32.5% retrieval failures and 31.0% generation failures—and 36.5% to architectural failures, including SEA errors at 24.5%; on multi-hop datasets, SEA errors rise to approximately 28–32%, making the audit mechanism itself a major bottleneck (asl et al., 25 Oct 2025). AT-RAG identifies sensitivity to topic assignment quality as a core limitation: incorrect or overly broad BERTopic assignments can reduce retrieval precision and answer accuracy (Rezaei et al., 2024).

In interactive database QA, brittleness is especially visible. EHR-ChatQA reports consistent AdaptQA failures dominated by value linking errors at 46.9%, SQL generation errors at 25.0%, rule violations at 15.6%, and limited clinical knowledge at 12.5%; inconsistent failures across trials are even more concentrated in SQL generation issues, at 71.8% (Lee et al., 27 Sep 2025). Task-Adaptive Embedding Refinement highlights another trade-off: if the teacher LLM’s judgments are biased or inaccurate, the refined embedding can inherit those errors, and with small $t$9 the updated query can overfit to idiosyncrasies in the feedback set $E(t) = E(t-1) \cup Filter(D_t, Q)$0 (Gera et al., 12 May 2026). IntentionReasoner notes that BU–BH distinctions remain difficult, that overly aggressive safety constraints can reduce informativeness, and that reinforcement learning slightly increases jailbreak ASR for the 1.5B and 3B variants relative to SFT-only models, even though RL improves rewriting utility and over-refusal behavior (Shen et al., 27 Aug 2025).

A common misconception is to equate AdaptQA exclusively with iterative retrieval loops. The literature includes strategy routing by complexity class, outcome-supervised Best-of-$E(t) = E(t-1) \cup Filter(D_t, Q)$1 reformulation, SQL and SPARQL repair under execution feedback, safety rewriting, and embedding-space refinement, so the “adaptive” component may apply before retrieval, during retrieval, after execution, or even at the representation level (Jeong et al., 2024, 2506.01381, Li et al., 2024, Jian et al., 3 Nov 2025, Gera et al., 12 May 2026). Another misconception is that more refinement is always better. Multiple studies report diminishing returns, quality regressions under excessive iteration, or length-control problems that required explicit reward terms or hard iteration caps.

6. Relation to adjacent paradigms and likely research directions

AdaptQA sits at the intersection of query reformulation, retrieval control, reinforcement learning, and interactive information access. AQA shows one path by casting adaptive QA as a contextual multi-armed bandit over communication graphs and rewards that combine performance with cost; its proposed extensions include richer context features, hierarchical bandits, and safety-aware rewards (Hoveyda et al., 2024). DIVA offers a different route: instead of fixed refinement loops, it first diversifies possible interpretations, then verifies coverage, and only then decides whether to proceed with RAG or closed-book generation; the paper explicitly suggests extending the verify signal into additional refinement rounds for low-coverage cases (In et al., 2024).

Two other strands broaden the notion of what is being refined. Entity-Centric Query Refinement formulates refinement selection as partitioning the answer entity set using the QRESP objective $E(t) = E(t-1) \cup Filter(D_t, Q)$2, thereby emphasizing coverage, non-redundancy, and balance rather than answer generation alone (Wadden et al., 2022). QREFINE, by contrast, treats question refinement as a Seq2Seq reinforcement-learning problem and optimizes both wording reward and answer correlation reward; the method improves readability and retrieval accuracy but also exposes the long-standing instability of reward-driven query generation (Liu et al., 2019). Automatic in-domain exemplar construction and multi-LLM refinement for query expansion adds a domain-adaptive query expansion view in which refinement is mediated by exemplar harvesting, clustering, and LLM consolidation rather than by direct answer supervision (Li et al., 9 Feb 2026).

The research directions explicitly proposed across these works are convergent. They include adaptive topic assigners in AT-RAG, stronger or learned verifiers in DIVA, stage-wise or hierarchical selectors in contextual-bandit frameworks, richer intent taxonomies and reward models in safeguard systems, more robust state representations for goal-shifting EHR agents, and active selection of informative feedback documents in embedding refinement (Rezaei et al., 2024, In et al., 2024, Hoveyda et al., 2024, Shen et al., 27 Aug 2025, Lee et al., 27 Sep 2025, Gera et al., 12 May 2026). This suggests that future AdaptQA systems will be less monolithic and more explicitly instrumented: they will expose internal state such as gap sets, difficulty labels, refinement candidates, or rewritten intent, and they will use those states not only to improve accuracy but also to bound cost, diagnose brittleness, and support human oversight.