What Makes a Good Query? Measuring the Impact of Human-Confusing Linguistic Features on LLM Performance

William Watson; Nicole Cho; Sumitra Ganesh; Manuela Veloso

cs.CLFeb 23, 2026

What Makes a Good Query? Measuring the Impact of Human-Confusing Linguistic Features on LLM Performance

Authors: William Watson, Nicole Cho, Sumitra Ganesh, Manuela Veloso

Organizations: J.P. Morgan AI Research

Abstract

Large Language Model (LLM) hallucinations are usually treated as defects of the model or its decoding strategy. Drawing on classical linguistics, we argue that a query's form can also shape a listener's (and model's) response. We operationalize this insight by constructing a 22-dimension query feature vector covering clause complexity, lexical rarity, and anaphora, negation, answerability, and intention grounding, all known to affect human comprehension. Using 369,837 real-world queries, we ask: Are there certain types of queries that make hallucination more likely? A large-scale analysis reveals a consistent "risk landscape": certain features such as deep clause nesting and underspecification align with higher hallucination propensity. In contrast, clear intention grounding and answerability align with lower hallucination rates. Others, including domain specificity, show mixed, dataset- and model-dependent effects. Thus, these findings establish an empirically observable query-feature representation correlated with hallucination risk, paving the way for guided query rewriting and future intervention studies.

Figures & tables

Figure 1: Prevalence of binary linguistic features across hallucination risk categories ( Safe , Borderline , Risky ). Warmer colors indicate higher frequency. Lack of specificity , clause complexity , and polysemous words show a pronounced rise from Safe to Risky .

	Ordinal ( $\beta$ -only)					Correlation
Feature	Coef	SE	$z$ -value	$p$ -value	OR	$\rho$	$\tau$	Adj. $p$	$p<0.05$
Lack of Specificity	0.868	0.010	85.898	<10 -5	2.382	0.271	0.256	<10 -5	✓
Clause Complexity	0.568	0.010	57.363	<10 -5	1.764	0.155	0.147	<10 -5	✓
Negation Usage	0.311	0.016	19.499	<10 -5	1.364	0.028	0.026	<10 -5	✓
Excessive Details	0.247	0.026	9.668	<10 -5	1.281	0.066	0.063	<10 -5	✓
Anaphora Usage	0.214	0.009	23.827	<10 -5	1.238	0.107	0.101	<10 -5	✓

Table 1: Results for Observed Risk analyses. Left: Ordinal logistic regression estimates (using both binary and scaled numeric predictors. Right: Spearman’s

\rho

and Kendall’s

\tau

correlation coefficients between each feature and Observed Risk, with adjusted

p

-values and a significance indicator. Features in italics (e.g., Lack of Specificity , Clause Complexity , Query Token Length , Number of Clauses , and Answerability ) highlight particularly intriguing effects. All adjusted

p

-values were below

10^{-5}

except for “Named Entities Present” (

p=0.115

, not significant).

Figure 2: ECDFs of predicted $P(\text{Risky})$ for Present vs. Absent (top six by KS). Shaded regions indicate dominance; inset shows KS and $\Delta$ median. Lack of Specificity, Excessive Details, Clause Complexity, and Query–Scenario Mismatch shift mass toward higher risk; Answerability and Intention Grounding shift mass lower.

Figure 3: Risk vs. query length by scenario. Each curve shows the empirical probability of a risky output (fraction of "Risky" labels) after quantile-binning query length within a scenario ( $\geq 50$ examples per bin). Risk rises with length for Abstractive , remains low/flat for Extractive , and is intermediate for Multiple-Choice . Takeaway: longer, open-ended queries are more hallucination-prone, while extractive settings remain robust across lengths.

Figure 4: Feature coefficients $\beta$ (left) and dataset/scenario fixed effects $\alpha,\gamma$ (right) from the ordinal logit model. Positive values increase log-odds of Risky . Answerability is strongly protective; Lack of Specificity , Negation , and Anaphora increase risk.

Figure 5: LODO coefficient stability (ordinal logit). Each point is a feature coefficient estimated when one dataset is held out (color = held-out dataset’s scenario), with short horizontal bars showing the mean and $\pm$ 1 s.d. across LODO runs. The blue diamond is the pooled (full-fit) coefficient. Signs and magnitudes are stable: Lack of Specificity , Clause Complexity , Query–Scenario Mismatch remain risk-increasing, while Answerability remains strongly protective.

Figure 6: Clustered Spearman correlation matrix using complete linkage & correlation distance. The color scale ranges from red (

\rho=1

, strong positive correlation) to blue (

\rho=-1

, strong negative correlation). Dendrograms group features with similar correlation patterns and share similar linguistic functions.

Appendix figures & tables16 assets

Supplementary material from the paper’s appendix.

Appendix

Query Type	Train	Val	Test	Total
Extractive	80,049	5,843	–	85,892
Multiple Choice	45,997	14,127	21,573	81,697
Abstractive	176,446	24,521	1,281	202,248
Overall	302,492	44,491	22,854	369,837

Appendix

Table 2: Number of queries across Extractive , Multiple Choice , and Abstractive categories, split by train, validation (Val), and test sets. Note that we make no distinction between these splits in our analysis.

	Safe		Borderline		Risky
Query Type	Count	%	Count	%	Count	%
Extractive	58,834	69.0	19,618	23.0	6,773	8.0
Multiple Choice	38,869	47.0	24,711	29.9	19,064	23.1
Abstractive	67,078	33.4	44,244	22.0	89,429	44.5

Appendix

Table 3: Observed Risk counts and row-normalized percentages across query types. Each risk group shows the count and percentage of predictions labeled as Safe , Borderline , or Risky .

Figure 7: Illustration of our query-based linguistic features, grouped by correlation categories. The ( + ) or ( - ) signs indicate whether each feature is positively or negatively associated with hallucination risk, while “ ± ” and “ O ” denote mixed or non-significant effects, respectively. Note: several syntactic features (e.g., token length, number of clauses) show negative associations , indicating richer structure can co-occur with lower risk.

Dataset	Scenario	Domain	License	Count	Citation
SQuADv2	E, A	Wikipedia	CC BY-SA 4.0	86K	Rajpurkar et al. (2016) ; Rajpurkar et al. (2018)
TruthfulQA	M, A	General Knowledge	Apache-2.0	807	Lin et al. (2022)
SciQ	M, A	Science	CC BY-NC 3.0	13K	Welbl et al. (2017)
MMLU	M	Various	MIT	15K	Hendrycks et al. (2021)
PIQA	M	Physical Commonsense	AFL-3.0	17K	Bisk et al. (2020)
BoolQ	M	Yes/No Questions	CC BY-SA 3.0	13K	Clark et al. (2019) ; Wang et al. (2019)

Appendix

Table 4: Overview of datasets used in our study, including domain, license, number of examples, and associated scenario types. These datasets span a diverse range of question types, knowledge areas, and reasoning skills, supporting robust evaluation across domains. Scenario types tested: E = Extractive, M = Multiple Choice, A = Abstractive.

Figure 8: ROC–AUC Landscape over Reward-Weight Simplex. Each point represents a convex combination of weights $(\alpha,\beta,\gamma)$ over the LLM-judge, Fuzzy, and BLEU metrics. Color indicates ROC–AUC measured on a held-out validation set; the shaded region denotes the top 1% frontier. Our selected weights $(0.6,0.3,0.1)$ are marked in red.

Figure 9: All features: ECDFs of predicted $P(\text{Risky})$ by feature presence. Same rendering as Figure 2 .

Feature	KS	$\Delta$ median	$n_{\text{abs}}$	$n_{\text{pres}}$	Direction
Answerability	$0.72$	$-0.58$	25,280	343,340	risk $\downarrow$
Intention Grounding	$0.66$	$-0.59$	13,576	355,044	risk $\downarrow$
Lack of Specificity	$0.56$	$0.42$	302,781	65,839	risk $\uparrow$
Excessive Details	$0.43$	$0.30$	361,479	7,141	risk $\uparrow$
Clause Complexity	$0.40$	$0.27$	307,849	60,771	risk $\uparrow$
Query–Scenario Mismatch	$0.40$	$0.34$	333,939	34,681	risk $\uparrow$

Appendix

Table 5: Feature ranking by ECDF separation. KS and

\Delta

median computed on predicted

P(\text{Risky})

Feature	$n_{\text{Present}}$	$n_{\text{Absent}}$	Overlap	ATE (IPW)	ATE (Strat.)
Named Entities Present	260,289	108,331	1.000	-0.000	-0.001
Polysemous Words	141,646	226,974	1.000	+0.014	+0.016
Pragmatic Features	97,727	270,893	0.993	+0.007	+0.006
Contextual Constraints	242,475	126,145	0.983	+0.000	+0.003
Domain Specificity	296,997	71,623	0.979	+0.001	+0.006
Clause Complexity	60,771	307,849	0.969	+0.103	+0.083

Appendix

Table 6: Propensity overlap and overlap-conditioned uplifts by feature. Overlap is the common-support share in

\pi_{f}(z)

(Present vs. Absent). ATEs are percentage-point changes in

\Pr(\textsc{Risky})

under IPW and propensity-stratified matching, reported only where overlap is adequate; bold marks salient non-zero effects. “—” indicates poor overlap (no uplift reported).

Figure 10: Risk vs. query length (Present vs Absent). For each feature, we bin queries into length quantiles and plot the empirical probability of the Risky label within each bin for Present vs Absent . Shaded bands are binomial CIs. Separation is largest at short lengths: features such as Lack of Specificity and Excessive Details increase risk, whereas Answerability and Intention Grounding reduce risk across lengths.

Figure 11: Heatmap of the percentage of queries exhibiting each binary linguistic feature, grouped by query type (extractive, multiple choice, abstractive) and categorized by observed risk level ( Safe , Borderline , Risky ). Warmer colors (reds) indicate higher prevalence of a feature, while cooler colors (blues) indicate lower prevalence. Several features ( lack of specificity , clause complexity , polysemous words ) increase most prominently from Safe to Risky , showing a clear monotonic rise in prevalence across risk categories. In contrast, answerability and intention grounding decrease steadily, and certain features ( domain specificity and contextual constraints ) display opposite trends across different query types.

Figure 12: Reliability by feature. Binned reliability curves for predicted $P(\text{Risky})$ when a feature is Present vs Absent . Dashed line is perfect calibration. An overall ECE is reported per panel (10 equal-mass bins). Calibration across strata. We show reliability curves stratified by feature presence. The model is reasonably calibrated across strata (ECE $\approx$ 0.05–0.06). Importantly, the direction of miscalibration does not reverse between Present/Absent strata for the dominant features (e.g., Answerability, Lack of Specificity), supporting that the feature effects observed in the ECDFs translate to well-behaved risk scores rather than artifacts of calibration.

Figure 13: Propensity overlap by feature. For each feature $f$ , we fit a logistic model $\pi_{f}(z)=\Pr(T_{f}{=}1\mid Z_{f}{=}z)$ over $Z_{f}=(x_{-f},\alpha,\gamma)$ and plot KDEs of $\hat{\pi}_{f}$ for Present ( $T_{f}{=}1$ ) vs. Absent ( $T_{f}{=}0$ ). Substantial overlap indicates adequate support for balancing or matching; near-degenerate propensities (mass near 0 or 1) warn that causal comparisons will be fragile. Several features (e.g., Answerability , Intention Grounding ) show limited overlap, which we treat with weighting and sensitivity analyses.

Panel of Figure 14 — Figure 14: Per-class probability KDEs. KDEs of model-predicted $P(\text{Safe})$ , $P(\text{Borderline})$ , and $P(\text{Risky})$ grouped by empirical class labels.

	Feature	Definition	Example
Structural	Query and Context Length	Total number of tokens in the query (and context, when applicable)	"How does reinforcement learning work?" (6 tokens)
	Anaphoric Reference	Presence of pronouns or references requiring external context.	"What about that one?" (Unclear reference)
	Clause Complexity	Measures the presence of multiple subordinate clauses	"While I was walking home, I saw a cat that looked just like my friend’s."
	Dependency Tree Depth	Depth of the query’s syntactic dependency tree.	"Describe the structure of a sentence that contains multiple levels of embedding." (Depth: 7)
	Parse Tree Height	Height of the parse tree, providing a secondary measure of syntactic complexity.	"Analyze a sentence with nested relative clauses." (Height: 4)
Scenario-Based	Query Type	Extractive, Multiple Choice, or Abstractive.	"Summarize the latest economic report." (Requires retrieval)

Appendix

Table 7: Summary of our feature categories, definitions, and examples (See Section 3 )

Feature	Presence	Question	Chain of Thought
Anaphora Usage	✓	Who was the guitarist for the English Rock band who Terry Kirkbride performed live in the studio with?	The question contains an anaphoric reference (‘the English Rock band’) without clear contextual information.
	✗	Isotopes are named for their number of protons plus what?	The question does not contain anaphoric references; it is a straightforward scientific inquiry.
Clause Complexity	✓	During evolution, something happened to increase the size of what organ in humans, relative to that of the chimpanzee?	The query has multiple clauses, increasing its complexity.
	✗	What do some animals do to adjust to hot temperatures?	The question is simple, consisting of a single clause.
Query-Scenario Mismatch	✓	What type of forested areas can be found on the highest terrace?	The query asks about ‘forested areas,’ but without a specific location or context, creating a mismatch.
	✗	What date in 2009 saw the heaviest UK snowfall since 1991?	The question has a direct and valid scenario, asking for a factual historical date.

Appendix

Table 8: Representative queries illustrating the presence and absence of selected linguistic features, with accompanying chain-of-thought explanations (Part 1).

Feature	Presence	Question	Chain of Thought
Polysemous Words	✓	Who supervised the sting operation that implicated Evelyn Dawn Knight?	The word ‘supervised’ could have different meanings but in this context refers to oversight.
	✗	Which string instrument often played the basso continuo parts?	The terms ‘string instrument’ and ‘basso continuo’ are not polysemous in this context.
Subjectivity	✓	What is a criticism of other streaming services?	The query invites subjective responses based on personal opinions.
	✗	What is the second book in the Harry Potter series?	The question is factual and does not involve subjectivity.
Answerability	✓	How long was Warsaw occupied by Germany?	The question can be answered based on explicit historical data.
	✗	Beyoncé would take a break from music in which year?	The event may not have a definitive, verifiable answer.

Appendix

Table 9: Representative queries illustrating the presence and absence of selected linguistic features, with accompanying chain-of-thought explanations (Part 2).

Hallucination in Language Models

Hallucination in language models studies how systems produce unsupported or false claims and how to detect, explain, and reduce them.

See all 254 papers

When the Wrong Key Wins: Understanding and Detecting Hallucinations in LLMs
Sep 14, 2026Xuhan Tong, Haoyue Bai, Dawei Zhou +2Language Model Error Detection Hallucination in Language Models
University of Wisconsin–Madison
A Geometric Analysis of Small-sized Language Model Hallucinations
Feb 16, 2026Emanuele Ricco, Elia Onofri, Lorenzo Cima +2Hallucination in Language Models LLM Hallucination Detection
Computer, Electrical and Mathematical Sciences and Engineering (CEMSE) division, King Abdullah University of Science and Technology (KAUST), Thuwal, Saudi Arabia · Istituto di Informatica e Telematica (IIT), National Research Council of Italy (CNR), Pisa, Italy · Department of Information Engineering, University of Pisa, Pisa, Italy
How Much Do LLMs Hallucinate in Document Q&A Scenarios? A 172-Billion-Token Study Across Temperatures, Context Lengths, and Hardware Platforms
Mar 9, 2026JV RoigHallucination in Language Models LLM Evaluation
Kamiwaza AI

Sep 14, 2026J/K move · Enter open · S save

When the Wrong Key Wins: Understanding and Detecting Hallucinations in LLMs

Xuhan Tong, Haoyue Bai, Dawei Zhou, Naichen Shi+1

University of Wisconsin–Madison

Language Model Error Detection Hallucination in Language Models LLM Reliability LLM Hallucination Detection Language Model Probing

Open paper AskPDF arXiv

Large language models can hallucinate even when the knowledge required for a correct answer is already available. We study this failure through a latent-key view of inference, where answer selection depends on competition among associations acquired during pretraining. We show that model predictions can be highly sensitive to individual query keywords, that these influential keywords exhibit entity-specific binding, and that their effects are systematically shaped by pretraining frequency. Multiple bindings can also compete and exhibit higher-order interactions within the same query. Based on this mechanism, we introduce a two-stage keyword-perturbation method for hallucination detection. By removing influential keywords and measuring how the model reorganizes its prediction, the method distinguishes errors caused by misleading key associations from correct decisions supported by diagnostic evidence. Across multiple models and benchmarks, perturbation provides a strong and transferable detection signal, reaching

0.910

AUROC on probe-known ScientistQA. Finally, we extend the same probabilistic framework to four hallucination regimes: knowledge deficit, wrong knowledge, context distraction, and unstable inference. Their operational distributions across benchmarks provide diagnostic context for why different detector families succeed in different settings.

Abstract

Hallucination in Language Models

Explore similar work

When the Wrong Key Wins: Understanding and Detecting Hallucinations in LLMs

A Geometric Analysis of Small-sized Language Model Hallucinations

How Much Do LLMs Hallucinate in Document Q&A Scenarios? A 172-Billion-Token Study Across Temperatures, Context Lengths, and Hardware Platforms