How You Ask Matters! Adaptive RAG Robustness to Query Variations

Yunah Jang; Megha Sundriyal; Kyomin Jung; Meeyoung Cha

cs.CLApr 12, 2026

How You Ask Matters! Adaptive RAG Robustness to Query Variations

Authors: Yunah Jang, Megha Sundriyal, Kyomin Jung, Meeyoung Cha

Organizations: Seoul National University · Max Planck Institute for Security and Privacy, Germany

Abstract

Adaptive Retrieval-Augmented Generation (RAG) promises accuracy and efficiency by dynamically triggering retrieval only when needed and is widely used in practice. However, real-world queries vary in surface form even with the same intent, and their impact on Adaptive RAG remains under-explored. We introduce the first large-scale benchmark of diverse yet semantically identical query variations, combining human-written and model-generated rewrites. Our benchmark facilitates a systematic evaluation of Adaptive RAG robustness by examining its key components across three dimensions: answer quality, computational cost, and retrieval decisions. We discover a critical robustness gap, where small surface-level changes in queries dramatically alter retrieval behavior and accuracy. Although larger models show better performance, robustness does not improve accordingly. These findings reveal that Adaptive RAG methods are highly vulnerable to query variations that preserve identical semantics, exposing a critical robustness challenge.

Figures & tables

Figure 1: Retrieval-decision flip rates of the Qwen-32B model, measuring how often the model changes its retrieval judgment under meaning-preserving human query rewrites. Higher rates signal instability. Flips are categorized as One-flip (only one rewrite flips) and Both-flip (both rewrites flip).

Figure 2: Example of Adaptive RAG responses under human, original, and LLM-generated query rewrites, highlighting differences in answer correctness, computation overhead, and retrieval score.

Original Query: Who is the child of the director of film Mukhyamantri (1996 Film)?
Category	Variation	Sample Rewrites
Style	Formality (Form.) $\downarrow$	Yo, quick one — who’s the kid of the director behind the 1996 flick Mukhyamantri?
	Readability (Read.) $\downarrow$	By whom is the progeny of the director of Mukhyamantri (1996 Film) constituted?
Type	Declarative (Decl.)	I want to know who is the child of the director of the film Mukhyamantri (1996 Film).
	Imperative (Impr.)	Name the child of the director of film Mukhyamantri (1996 Film).
Error	Spelling (Spell.) $\downarrow$	Who is ght childer ar the directer of filme Mukhyamantri (1996 Film )?

Table 1: Examples of query rewrites grouped by category, illustrating stylistic, structural, and error-based variations.

\downarrow

denotes reduction in corresponding writing variation.

Panel of Figure 3 — Figure 3: InAccuracy results across six perturbation types for a given dataset / model pair.

			SQuAD							2WIKI
	Target	Mtd.	RE $\downarrow$						CVR $\uparrow$	RE $\downarrow$						CVR $\uparrow$
		Mtd.	Form.	Read.	Decl.	Impr.	Spell.	Gram.	CVR $\uparrow$	Form.	Read.	Decl.	Impr.	Spell.	Gram.	CVR $\uparrow$
Llama-3.1	RTR	LB	0.472	0.592	0.416	0.338	0.431	0.439	0.734	0.853	0.859	0.455	0.846	0.928	0.941	0.710
		GB	0.092	0.094	0.086	0.090	0.100	0.092	0.937	0.468	0.416	0.402	0.542	0.496	0.435	0.742
	LLM	LB	0.441	0.592	0.396	0.313	0.416	0.411	0.767	1.001	1.001	0.427	1.007	1.093	1.130	0.742
		GB	0.056	0.059	0.048	0.055	0.061	0.051	0.963	0.447	0.404	0.401	0.520	0.477	0.418	0.745

Table 3: Computation robustness across different model-generated query variations. Darker shading indicates more desirable outcomes: lower RE and higher CVR . We compare the results between logit-based (LB) and generation-based (GB) method. RTR stands for retriever and bold denotes maximum across the variations.

Model	Mtd.	Orig.	Human	Spell.
Llama-3.1-8B	LB	37.6	17.9	23.0
Llama-3.1-8B	GB	20.2	12.1	9.8
QwQ-32B	LB	26.6	20.0	21.0
QwQ-32B	GB	59.2	43.4	45.8

Table 4: Contriever (dense retriever) results. We report accuracy only for Original, Human, and Spelling variation under logit-based (LB) and generation-based (GB) methods.

Figure 4: Under- and over-confident rates on model query variations.

Figure 5: Query analysis. From left to right: (a) change in LM loss relative to the original query (

\Delta

Loss; nats/token), (b) change in query length (

\Delta

Length; words), and (c) semantic similarity to the original query (cosine similarity).

Panel of Figure 6 — Figure 6: Subquery drift effect on retrieval quality. 2WIKI Dragin results (a) Subquery similarity to the original decreases across turns. (b) Retrieval failure rate (no gold document in top- $k$ ) across turns.

Panel of Figure 7 — Figure 7: InAccuracy results with human rewrites.

	Original Query	Human Query
	Which film has the director born earlier, The Adventures Of Priscilla, Queen Of The Desert or Harvest: 3,000 Years?	Who was born first director of The Adventures Of Priscilla, Queen Of The Desert or Harvest: 3,000 Yearst
t	Original Subqueries	Human Subqueries	Sim	Hit(O)	Hit(H)
1	director of the adventures of priscilla, queen of the desert	director of the adventures of priscilla queen of the desert	0.99	✓	✓
2	director of harvest 3000 years	director of harvest 3000 years	1.00	✓	✓
3	“harvest: 3,000 years” director birth year	harvest 3000 years movie director	0.85	✓	✗
4	giuliano carnimeo birth year	harvest 3000 years korean movie director	0.13	✓	✗

Table 5: Per-hop comparison of original and human subqueries. Sim is the semantic similarity between the two subqueries at hop

t

. Hit(O) indicates whether the retrieved top-

k

set contains any ground-truth supporting evidence for the original (O) and human (H) query.

Figure 8: Gold document inclusion rate across hops. For each hop, we measure how often the top-

k

retrieved results contain at least one gold document.

Appendix figures & tables32 assets

Supplementary material from the paper’s appendix.

Appendix

Panel of Table 7 — Table 7: Prompt for lower-readability query generation.

Dataset	Mtd.	Original	Human	Form.	Read.	Decl.	Impr.	Spell.	Gram.
Llama-3.1-8B
2WIKI	LB	37.6 / 2.16	17.9 / 2.01	23 / 2.03	35 / 2.11	34 / 2.11	36 / 2.04	23.0 / 2.23	35.6 / 2.14
2WIKI	GB	20.2 / 8.24	12.1 / 7.27	19 / 7.15	18.6 / 7.41	19.4 / 7.36	20.4 / 7.46	9.8 / 6.53	21.2 / 7.03
SQuAD	LB	27.6 / 1.69	14.5 / 1.82	19.4 / 1.80	25 / 1.81	25.4 / 1.77	27.0 / 1.81	17.6 / 1.77	24.4 / 1.80
SQuAD	GB	16.4 / 1.93	11.0 / 1.95	15.4 / 4.90	16.2 / 4.90	16.8 / 4.87	15.8 / 3.92	14.8 / 1.94	13.8 / 1.92
QwQ-32B

Appendix

Table 8: Results for Dense Retriever (Contriever). We report results as (accuracy / retrieval count), comparing results between NR (NR), logit-based (LB), and generation-based (GB) method.

Temp 0.0	Temp 0.3	Temp 0.7	Temp 1.0
44.60%	45.33%	45.45%	42.67%

Appendix

Table 9: Flip rates (%) across different decoding temperatures on SQuAD.

Model	Dataset	Original	Human
Llama-3.1-8B	2WIKI	20.0 / 7.34	14.2 / 7.09
Llama-3.1-8B	SQuAD	16.8 / 1.92	10.5 / 1.95
QwQ-32B	2WIKI	61.0 / 4.32	57.1 / 4.06
QwQ-32B	SQuAD	33.0 / 1.40	28.5 / 1.42
GPT-4o mini	2WIKI	27.0 / 3.77	14.6 / 4.95
GPT-4o mini	SQuAD	17.0 / 2.34	16.0 / 2.34

Appendix

Table 10: Closed- and open-source model results We report results as (accuracy / retrieval count).

Dataset	Variation	F1	InAccuracy
2WIKI	Original	0.396	37.0
2WIKI	Human	0.283	23.8
SQuAD	Original	0.352	29.2
SQuAD	Human	0.260	17.0

Appendix

Table 11: Comparison between InAccuracy and token-level F1 scores We report the scores for logit-based method with Llama-3.1-8B across datasets and query variations.

Figure 9: Helpfulness and harmfulness distribution on NQ using the QwQ-32B model.

Dataset	Mtd.	Orig.	Human	Form.	Read.	Decl.	Impr.	Spell.	Gram.
Llama-3.1-8B
SQuAD	SR	$\mathbf{30.1}{\scriptstyle\pm 1.1}$	$19.3{\scriptstyle\pm 1.0}$	$22.2{\scriptstyle\pm 0.9}$	$26.3{\scriptstyle\pm 0.7}$	$26.6{\scriptstyle\pm 0.5}$	$29.3{\scriptstyle\pm 0.9}$	$18.8{\scriptstyle\pm 1.0}$	$26.7{\scriptstyle\pm 0.9}$
	LB	$\mathbf{28.8}{\scriptstyle\pm 0.5}$	$17.2{\scriptstyle\pm 0.3}$	$22.2{\scriptstyle\pm 0.6}$	$26.3{\scriptstyle\pm 2.0}$	$26.9{\scriptstyle\pm 0.7}$	$28.3{\scriptstyle\pm 2.1}$	$21.2{\scriptstyle\pm 0.7}$	$25.9{\scriptstyle\pm 0.9}$
	GB	$15.7{\scriptstyle\pm 1.8}$	$9.8{\scriptstyle\pm 1.6}$	$\mathbf{16.0}{\scriptstyle\pm 2.0}$	$15.1{\scriptstyle\pm 1.7}$	$15.8{\scriptstyle\pm 2.0}$	$14.8{\scriptstyle\pm 0.5}$	$14.1{\scriptstyle\pm 1.8}$	$15.3{\scriptstyle\pm 0.9}$
NQ	SR	$35.3{\scriptstyle\pm 0.1}$	$28.0{\scriptstyle\pm 0.0}$	$\mathbf{40.1}{\scriptstyle\pm 1.3}$	$35.2{\scriptstyle\pm 0.6}$	$38.8{\scriptstyle\pm 1.1}$	$37.1{\scriptstyle\pm 1.8}$	$28.1{\scriptstyle\pm 1.0}$	$37.2{\scriptstyle\pm 0.0}$
	LB	$41.8{\scriptstyle\pm 0.3}$	$33.6{\scriptstyle\pm 0.6}$	$44.7{\scriptstyle\pm 1.8}$	$43.8{\scriptstyle\pm 0.6}$	$46.1{\scriptstyle\pm 0.8}$	$\mathbf{47.2}{\scriptstyle\pm 1.2}$	$35.6{\scriptstyle\pm 0.3}$	$44.9{\scriptstyle\pm 1.6}$

Appendix

Table 13: Multi-run results for Llama-3.1-8B (mean

\pm

std over 3 runs). Methods: NR=Naive Retrieval, LB=Logit-based, GB=Generation-based. For each row, the best-performing result is shown in bold and the second-best result is underlined.

Dataset	Mtd.	Orig.	Human	Form.	Read.	Decl.	Impr.	Spell.	Gram.
QwQ-32B
SQuAD	NR	$\mathbf{28.8}{\scriptstyle\pm 0.3}$	$18.3{\scriptstyle\pm 0.8}$	$22.5{\scriptstyle\pm 0.4}$	$18.1{\scriptstyle\pm 1.0}$	$19.0{\scriptstyle\pm 1.1}$	$17.7{\scriptstyle\pm 1.3}$	$17.1{\scriptstyle\pm 2.5}$	$18.7{\scriptstyle\pm 1.0}$
	LB	$19.0{\scriptstyle\pm 0.3}$	$15.8{\scriptstyle\pm 2.1}$	$25.4{\scriptstyle\pm 0.7}$	$18.0{\scriptstyle\pm 0.3}$	$\mathbf{28.4}{\scriptstyle\pm 1.1}$	$18.2{\scriptstyle\pm 1.4}$	$16.2{\scriptstyle\pm 1.1}$	$18.8{\scriptstyle\pm 0.3}$
	GB	$\mathbf{38.2}{\scriptstyle\pm 0.7}$	$28.8{\scriptstyle\pm 1.5}$	$29.5{\scriptstyle\pm 0.5}$	$32.1{\scriptstyle\pm 1.6}$	$33.2{\scriptstyle\pm 3.1}$	$31.6{\scriptstyle\pm 5.1}$	$29.2{\scriptstyle\pm 7.4}$	$32.1{\scriptstyle\pm 3.0}$
NQ	NR	$42.3{\scriptstyle\pm 1.2}$	$37.9{\scriptstyle\pm 1.4}$	$44.1{\scriptstyle\pm 0.8}$	$45.1{\scriptstyle\pm 1.9}$	$43.4{\scriptstyle\pm 0.3}$	$44.7{\scriptstyle\pm 0.8}$	$35.0{\scriptstyle\pm 1.8}$	$\mathbf{45.7}{\scriptstyle\pm 0.7}$
	LB	$33.1{\scriptstyle\pm 1.0}$	$28.7{\scriptstyle\pm 0.2}$	$35.8{\scriptstyle\pm 0.2}$	$36.0{\scriptstyle\pm 0.5}$	$\mathbf{37.5}{\scriptstyle\pm 1.0}$	$36.4{\scriptstyle\pm 1.2}$	$29.6{\scriptstyle\pm 1.7}$	$\mathbf{37.5}{\scriptstyle\pm 0.5}$

Appendix

Table 14: Multi-run results for QwQ-32B (mean

\pm

std over 3 runs). Methods: NR=Naive RAG, LB=Logit-based, GB=Generation-based.For each row, the best-performing result is shown in bold and the second-best result is underlined.

Human Annotation Instruction

Task Overview 1. You will be given a query (question) and its corresponding answer. 2. Your task is to rewrite the given query — imagine how you would ask the same question if you were speaking to a large language model (like ChatGPT). What You Should Do 1. Understand the original query.

\circ

Identify what the user is trying to find out.

\circ

Grasp the intent and main topic clearly. 2. Rewrite the query in a new way while keeping the same meaning.

\circ

You may change the tone, phrasing, or structure.

\circ

Information loss, addition, or simplification is acceptable.

\circ

The goal is to make the rewritten query similar in meaning but different in expression.

\circ

Make sure that your rewritten query can still be answered appropriately with the same answer as the original query. 3. Incorporate your personal style.

\circ

You can make the question sound more casual, detailed, concise, or clear.

\circ

Different writing styles and tones are encouraged. 4. Avoid copying the original query.

\circ

Do not simply rephrase it with only minor surface changes.

\circ

The rewritten query should feel genuinely reworded and natural. 5. Use the answer only for context.

\circ

The answer helps you understand what the original query intended to ask.

\circ

You should not change your rewrite based on the answer’s content.

\circ

Your task is to rewrite the question, not the answer.

Appendix

Table 15: Human Annotation Instruction: Query Rewriting Task.

Figure 10: InAccuracy (Llama-3.1-8B). Performance across datasets for four Adaptive RAG methods. Each panel reports InAccuracy on seven query variations: the original query and six model-generated perturbations.

Figure 11: InAccuracy (QwQ-32B). Performance across datasets for four Adaptive RAG methods. Each panel reports InAccuracy on seven query variations: the original query and six model-generated perturbations).

Figure 12: InAccuracy (Llama-3.1-8B) including human rewrites. Performance across datasets for four Adaptive RAG methods, evaluated on a matched subset of instances that have human rewrites. For each panel, InAccuracy is reported for eight query variations (original, human rewrite, and six model-generated perturbations), all computed on the same human-rewrite subset.

Figure 13: InAccuracy (QwQ-32B) including human rewrites. Performance across datasets for four Adaptive RAG methods, evaluated on a matched subset of instances that have human rewrites. For each panel, InAccuracy is reported for eight query variations (original, human rewrite, and six model-generated perturbations), all computed on the same human-rewrite subset.

Dataset	Method	Similarity $\uparrow$						Diversity $\downarrow$
		Form.	Read.	Decl.	Impr.	Spell.	Gram.	Diversity $\downarrow$
Llama-3.1-8B
SQuAD	WR	0.568	0.580	0.598	0.635	0.538	0.584	0.459
	NR	0.509	0.552	0.606	0.631	0.466	0.578	0.499
	LB	0.551	0.539	0.593	0.634	0.552	0.599	0.469
	GB	0.426	0.408	0.433	0.430	0.411	0.414	0.585

Appendix

Table 16: Answer similarity to the original query and diversity across perturbation types (excluding human rewrites). We compare w/o retrieval (WR), naive RAG (NR), logit-based (LB), and generation-based (GB). Darker shading indicates more desirable outcomes: higher Similarity and lower Diversity.

Dataset	Method	Similarity $\uparrow$						Diversity $\downarrow$
		Form.	Read.	Decl.	Impr.	Spell.	Gram.	Diversity $\downarrow$
QwQ-32B
SQuAD	WR	0.566	0.576	0.602	0.634	0.549	0.593	0.438
	NR	0.511	0.570	0.625	0.649	0.488	0.599	0.475
	LB	0.573	0.579	0.615	0.631	0.544	0.584	0.440
	GB	0.552	0.558	0.579	0.589	0.552	0.579	0.459

Appendix

Table 17: Answer similarity to the original query and diversity across perturbation types (excluding human rewrites). We compare w/o retrieval (WR), naive RAG (NR), logit-based (LB), and generation-based (GB). Darker shading indicates more desirable outcomes: higher Similarity and lower Diversity.

Dataset	Method	Similarity $\uparrow$							Diversity $\downarrow$
		Human	Form.	Read.	Decl.	Impr.	Spell.	Gram.	Diversity $\downarrow$
Llama-3.1-8B
SQuAD	WR	0.542	0.579	0.610	0.644	0.653	0.542	0.600	0.446
	NR	0.505	0.580	0.544	0.613	0.622	0.479	0.607	0.503
	LB	0.527	0.557	0.557	0.599	0.626	0.566	0.584	0.469
	GB	0.447	0.416	0.421	0.423	0.449	0.427	0.449	0.583

Appendix

Table 18: Answer similarity to the original query and diversity across perturbation types (including human rewrites). We compare w/o retrieval (WR), naive RAG (NR), logit-based (LB), and generation-based (GB). Darker shading indicates more desirable outcomes: higher Similarity and lower Diversity.

Dataset	Method	Similarity $\uparrow$							Diversity $\downarrow$
		Human	Form.	Read.	Decl.	Impr.	Spell.	Gram.	Diversity $\downarrow$
QwQ-32B
SQuAD	WR	0.585	0.567	0.608	0.630	0.642	0.583	0.605	0.448
	NR	0.556	0.514	0.619	0.655	0.664	0.518	0.584	0.459
	LB	0.565	0.608	0.590	0.608	0.635	0.577	0.599	0.435
	GB	0.544	0.549	0.560	0.596	0.601	0.575	0.588	0.461

Appendix

Table 19: Answer similarity to the original query and diversity across perturbation types (including human rewrites). We compare w/o retrieval (WR), naive RAG (NR), logit-based (LB), and generation-based (GB). Darker shading indicates more desirable outcomes: higher Similarity and lower Diversity.

Figure 14: LLM Call on Llama-3.1-8B on both model-generated and human queries.

Figure 15: LLM Call on QwQ-32B on both model-generated and human queries.

Figure 16: Retriever Call on Llama-3.1-8B on both model-generated and human queries.

Figure 17: Retriever Call on QwQ-32B on both model-generated and human queries.

Dataset	Target	Method	RE $\downarrow$						CVR $\uparrow$
			Form.	Read.	Decl.	Impr.	Spell.	Gram.	CVR $\uparrow$
Llama-3.1-8B
SQuAD	Retriever	LB	0.472	0.592	0.416	0.338	0.431	0.439	0.734
		GB	0.092	0.094	0.086	0.090	0.100	0.092	0.937
	LLM	LB	0.441	0.592	0.396	0.313	0.416	0.411	0.767
		GB	0.056	0.059	0.048	0.055	0.061	0.051	0.963

Appendix

Table 20: Computation robustness across query variations (Llama-3.1-8B; no human rewrites). Darker shading indicates more desirable outcomes: lower RE and higher CVR.

Dataset	Target	Method	RE $\downarrow$						CVR $\uparrow$
			Form.	Read.	Decl.	Impr.	Spell.	Gram.	CVR $\uparrow$
QwQ-32B
SQuAD	Retriever	LB	0.494	0.450	0.405	0.415	0.495	0.441	0.740
		GB	0.360	0.322	0.298	0.315	0.361	0.330	0.762
	LLM	LB	0.477	0.466	0.389	0.399	0.491	0.424	0.760
		GB	0.265	0.230	0.225	0.238	0.250	0.243	0.858

Appendix

Table 21: Computation robustness across query variations (QwQ-32B; no human rewrites). Darker shading indicates more desirable outcomes: lower RE and higher CVR.

Dataset	Target	Method	RE $\downarrow$							CVR $\uparrow$
			Human	Form.	Read.	Decl.	Impr.	Spell.	Gram.	CVR $\uparrow$
Llama-3.1-8B
SQuAD	Retriever	LB	0.415	0.452	0.577	0.423	0.344	0.460	0.450	0.748
		GB	0.105	0.090	0.100	0.105	0.090	0.120	0.120	0.946
	LLM	LB	0.407	0.418	0.587	0.398	0.306	0.414	0.414	0.778
		GB	0.061	0.055	0.060	0.065	0.048	0.068	0.067	0.969

Appendix

Table 22: Computation robustness across query variations (Llama-3.1-8B; including human rewrites). Darker shading indicates more desirable outcomes: lower RE and higher CVR.

Dataset	Target	Method	RE $\downarrow$							CVR $\uparrow$
			Human	Form.	Read.	Decl.	Impr.	Spell.	Gram.	CVR $\uparrow$
QwQ-32B
SQuAD	Retriever	LB	0.439	0.500	0.568	0.446	0.404	0.451	0.547	0.741
		GB	0.370	0.295	0.280	0.310	0.240	0.335	0.290	0.766
	LLM	LB	0.461	0.468	0.597	0.426	0.389	0.417	0.494	0.757
		GB	0.285	0.252	0.233	0.242	0.210	0.273	0.242	0.859

Appendix

Table 23: Computation robustness across query variations (QwQ-32B; including human rewrites). Darker shading indicates more desirable outcomes: lower RE and higher CVR.

Figure 18: Under- and Over-confidence results on Llama-3.1-8B on model-generated queries.

Figure 19: Under- and Over-confidence results on QwQ-32B on model-generated queries.

Figure 20: Under- and Over-confidence results on Llama-3.1-8B on including human-written queries.

Figure 21: Under- and Over-confidence results on Llama-3.1-8B on including human-written queries.

Figure 22: Loss score across all datasets

Figure 23: Query length across all datasets

Figure 24: Semantic similarity to original query across all datasets

Adaptive RAG

Adaptive retrieval-augmented generation (RAG) adjusts when, what, and how much evidence a language model retrieves for each query.

See all 59 papers

Can QPP Choose the Right Query Variant? Evaluating Query Variant Selection for RAG Pipelines
Apr 24, 2026Negar Arabzadeh, Andrew Drozdov, Michael Bendersky +1Retrieval-Augmented Generation RAG Evaluation
UC Berkeley · Berkeley, CA, United States · Databricks +1
Agent-Orchestrated Adaptive RAG: A Comparative Study on Structured and Multi-Hop Retrieval
Jun 4, 2026Anuj Maharjan, Devinder Kaur, Richard MolyetRetrieval-Augmented Generation Multi-Hop Retrieval
Dept. of Electrical Engineering and Computer Science University of Toledo Toledo, OH, USA
Retriever Portfolios: A Principled Approach to Adaptive RAG
May 29, 2026Miltiadis Stouras, Vincent Cohen-Addad, Silvio Lattanzi +1Retrieval-Augmented Generation Adaptive RAG
1EPFL · 2Google Research

Apr 24, 2026J/K move · Enter open · S save

Can QPP Choose the Right Query Variant? Evaluating Query Variant Selection for RAG Pipelines

Negar Arabzadeh, Andrew Drozdov, Michael Bendersky, Matei Zaharia

UC Berkeley · Berkeley, CA, United States · Databricks+1

Retrieval-Augmented Generation RAG Evaluation

Open paper AskPDF arXiv

Large Language Models (LLMs) have made query reformulation ubiquitous in modern retrieval and Retrieval-Augmented Generation (RAG) pipelines, enabling the generation of multiple semantically equivalent query variants. However, executing the full pipeline for every reformulation is computationally expensive, motivating selective execution: can we identify the best query variant before incurring downstream retrieval and generation costs? We investigate Query Performance Prediction (QPP) as a mechanism for variant selection across ad-hoc retrieval and end-to-end RAG. Unlike traditional QPP, which estimates query difficulty across topics, we study intra-topic discrimination - selecting the optimal reformulation among competing variants of the same information need. Through large-scale experiments on TREC-RAG using both sparse and dense retrievers, we evaluate pre- and post-retrieval predictors under correlation- and decision-based metrics. Our results reveal a systematic divergence between retrieval and generation objectives: variants that maximize ranking metrics such as nDCG often fail to produce the best generated answers, exposing a "utility gap" between retrieval relevance and generation fidelity. Nevertheless, QPP can reliably identify variants that improve end-to-end quality over the original query. Notably, lightweight pre-retrieval predictors frequently match or outperform more expensive post-retrieval methods, offering a latency-efficient approach to robust RAG.

Abstract

Adaptive RAG

Explore similar work

Can QPP Choose the Right Query Variant? Evaluating Query Variant Selection for RAG Pipelines

Agent-Orchestrated Adaptive RAG: A Comparative Study on Structured and Multi-Hop Retrieval

Retriever Portfolios: A Principled Approach to Adaptive RAG