Source-linked AI summary
What Makes a Good Query? Measuring the Impact of Human-Confusing Linguistic Features on LLM Performance
William Watson, Nicole Cho, Sumitra Ganesh, Manuela Veloso
TL;DR
The paper asks whether linguistic query form contributes to LLM hallucination risk, rather than treating hallucination solely as a model or decoding defect. It builds a linguistically grounded feature representation and analyzes 369,837 real-world queries, finding a consistent risk landscape in which underspecification and deep clause nesting align with higher risk while intention grounding and answerability align with lower risk.
Problem
The paper addresses the limited large-scale evidence linking human-confusing linguistic query features to LLM hallucination behavior.
Method
The study extracts a 17-dimensional query feature vector and models ordinal hallucination risk across 369,837 queries using semantics-preserving paraphrases and dataset/scenario covariates.
Results
Underspecification and deep clause nesting align with higher hallucination propensity, while clear intention grounding and answerability align with lower risk; domain specificity has mixed effects.
Takeaways & Limitations
Query-feature representations can support feature-aware triage and low-effort rewrites aimed at proactively addressing hallucination risk before generation.
Takeaways & Limitations
The study is observational, limited to English queries and one class of LLMs, and its feature associations should not be interpreted causally.
Abstract
from arXiv · showhide
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.
1 Introduction
The paper examines whether query form shapes LLM hallucination risk, extending proactive mitigation beyond post-generation inspection. It introduces a linguistically grounded feature representation and maps query features to hallucination propensity across 369,837 real-world queries.
- Prior work emphasizes reactive mitigation, while comparatively fewer studies examine query-side factors beyond ambiguity detection.
- The study defines a 17-dimensional query feature vector covering structural, lexical, stylistic, and semantic properties relevant to comprehension.
- 369,837 real-world queries spanning 13 QA datasets, three scenarios, and 16 configurations support the empirical analysis.
- Underspecification and deep clause nesting align with higher hallucination propensity, while clear intention grounding and answerability align with lower risk.
- Domain specificity shows mixed, dataset- and model-dependent effects, while word rarity, superlatives, and complex negation show minimal association.
- The resulting risk landscape motivates feature-aware triage and low-effort query rewrites before generation.
2 Related Work
Related work addresses hallucinations through both proactive input-time methods and reactive post-generation methods. More recent studies also examine how query structure and semantic properties shape LLM outputs.
- Proactive methods enrich context before decoding through approaches such as retrieval-augmented generation and external tool use.
- Reactive methods evaluate or rerank outputs after decoding using techniques such as self-consistency and logit-based detectors.
- Query structure and semantic properties, including polysemy, contextual nuance, and specificity, influence LLM outputs in prior studies.
- The paper positions its contribution within emerging pre-generation query evaluation research.
3 Methodology
The methodology represents queries with binary linguistic features, constructs observed risk from semantics-preserving paraphrases, and models ordinal risk while evaluating distributional separation, calibration, overlap, and dataset robustness.
- Problem setup: Each query receives a SAFE, BORDERLINE, or RISKY ordinal label, modeled from linguistic features and dataset/scenario covariates.
- Linguistic features: The feature set spans ambiguity, referential structure, complexity, polarity, grounding, and other linguistic dimensions.
- Observed risk via semantics-preserving perturbations: Six lexically perturbed paraphrases are retained when their hybrid similarity to the original query is at least 0.85.
- Observed risk via semantics-preserving perturbations: A perturbation is hallucinated when the convex proxy exceeds 0.5, producing Safe, Borderline, and Risky categories from 0/6, 1–3/6, and 4–6/6 hallucinated paraphrases.
- Metrics and diagnostics: The analysis combines ordinal logistic coefficients with ECDF separations, calibration curves, length interactions, and propensity-overlap diagnostics.
- Robustness: Leave-One-Dataset-Out refits preserve coefficient signs and relative magnitudes, indicating that the risk landscape is not driven by one dataset.
4 Experimental Setup
Experiments use a fixed GPT-4o configuration across 13 QA datasets, three scenarios, and 16 total configurations. The setup evaluates risk construction, feature extraction, ordinal modeling, and robustness analyses across varied query settings.
- All generations use gpt-4o-2024-08-06 with a fixed prompting recipe and temperature τ = 1.0 for answering and paraphrase sampling.
- The benchmark covers 13 QA datasets spanning Multiple Choice, Abstractive, and Extractive scenarios.
- Risk rises with query length in Abstractive settings, remains low and flat in Extractive settings, and is intermediate for Multiple Choice.
- 369,837 query–response pairs are analyzed, with dataset and scenario entering the ordinal model as fixed effects.
- Structured detectors assign binary feature labels and rationales, with rubrics calibrated on a 100-sample held-out set.
- Risk labels derive from the paraphrase-based convex hallucination proxy, and ordinal coefficients are checked against ECDF separations.
5 Results: A Query-Feature Risk Landscape for Hallucination
The analysis identifies a consistent hallucination risk landscape: ambiguity and some complexity features align with higher risk, while answerability and intention grounding align with lower risk. Absolute risk varies by dataset and scenario, but feature directions remain broadly stable.
- ECDF comparisons likewise place Lack of Specificity, Excessive Details, Clause Complexity, and Query–Scenario Mismatch toward higher risk, while Answerability and Intention Grounding shift risk lower.
- Abstractive settings show sharply rising risk with length, whereas Extractive settings remain comparatively flat; nevertheless, feature-effect directions remain stable across bins.
- Leave-one-dataset-out fits preserve the signs and relative magnitudes of dominant features, indicating that the risk landscape is not driven by one dataset.
- Lack of Specificity has the highest positive coefficient (0.868) and odds ratio (OR) of 2.382, associating omitted details or aims with higher-risk outputs.
- Answerability has the largest protective coefficient (-1.106, OR = 0.331), while Intention Grounding (-0.168) is also negatively associated with hallucination risk.
- Query Token Length (-0.212), Dependency Depth (-0.128), and Number of Clauses (-0.262) are inversely correlated with risk, potentially because richer structure provides helpful context.
- Negation Usage (0.311) and Anaphora Usage (0.214) are positively associated with hallucination risk, but their effects are weaker than those of ambiguity and structure features.
- Domain Specificity has a near-zero coefficient (0.003, ρ = -0.013), while Named Entities Present is not statistically significant (p = 0.205).
6 Conclusion
The paper interprets hallucination risk as related to how strongly a query commits the model to a determinate reading. Grounding features constrain interpretation, whereas underspecified or structurally intricate queries expand the space for speculative completion.
- Queries that declare intent and make answerability explicit constrain the hypothesis space, while underspecified or structurally intricate queries expand it.
Disclaimer
The document includes a JPMorgan informational-purpose disclaimer and disclaims liability for the information’s completeness, accuracy, or reliability.
- The document is informational and is not investment research, investment advice, or a recommendation, offer, or solicitation.
Limitations
The study is primarily observational and its feature correlations should be interpreted as empirical associations rather than causal effects. Its scope is also limited to English queries, one class of LLMs, and independently modeled features.
- The study provides quasi-causal evidence at best, despite overlap diagnostics and ablations used to qualify comparisons.Several features, including Answerability, are semantic and cannot be cleanly manipulated without changing meaning.
- The experiments cover English-language queries and one class of LLMs, excluding multimodal inputs and evolving model behavior across versions.
- Feature extraction may introduce parsing errors in noisy queries because it relies on existing NLP toolkits and LLM predictions.
- The analysis treats linguistic features as independent variables and does not model higher-order interactions.Future work could examine whether feature combinations jointly contribute to hallucination risk.
- The reward formulation partly relies on an LLM-based judge, which may introduce systematic biases despite Pareto-optimal ROC–AUC tuning.
A Distribution of Hallucination Across Query Type
Hallucination prevalence varies substantially by query scenario, while linguistic features provide additional signals associated with risk. Abstractive queries are most often hallucination-prone, whereas extractive queries are comparatively safe, and ambiguity-related features tend to increase risk.
- Query-type distributions: Extractive queries have infrequent hallucinations and are mostly classified as Safe, likely because they include explicit supporting context.
- Query-type distributions: Multiple-choice distractor options correspond with a higher proportion of Borderline cases.
- Query-type distributions: Abstractive queries are most frequently associated with hallucinations, with a large share labeled Risky.
- Structural features: The analysis measures clause complexity through subordinate-clause counts, dependency depth, and parse-tree height.
C Reward-Weight Simplex Analysis:
The reward-weight analysis tests mixtures of semantic, fuzzy-match, and lexical signals using ROC–AUC. The selected mix lies on a high-AUC Pareto plateau, while fuzzy-match and BLEU-heavy weighting degrade performance.
- The analysis evaluates reward weights over a triangular simplex and computes ROC–AUC on a 100-item human-labeled validation set.
- AUC shifts by <0.5% when w0 varies by ±0.2, indicating robustness to substantial changes in the LLM-judge weight.
- Small increases in w1 rapidly leave the Pareto region, so the fuzzy-match term requires careful tuning.
- AUC steadily declines as w2 increases and is lowest at w2 = 1, where BLEU-only weighting overemphasizes surface overlap over semantic correctness.
- The final weights are (0.6, 0.3, 0.1), selected within a high-AUC Pareto plateau as a trade-off among semantic, fuzzy, and lexical signals.
D Feature Calculation Methodology
The study extracts linguistic features with structured LLM outputs and parsers, then models ordinal hallucination risk and evaluates feature-level distributional and propensity-based contrasts.
- Feature extraction: A structured-output LLM assigns binary linguistic-feature labels using typed schemas, operational rubrics, and positive/negative examples.The feature set includes phenomena involving anaphora, clause complexity, negation, answerability, intention grounding, and contextual constraints.
- Feature extraction: spaCy computes syntactic measures including clause counts, dependency depth, and parse-tree height, while tiktoken measures token length.Observed risk is derived from responses generated with gpt-4o-2024-08-06.
- Ordinal risk modeling: An ordinal logistic regression estimates associations between binary linguistic features and the three-level outcome Safe, Borderline, or Risky.Positive coefficients indicate higher likelihood of hallucination risk, with dataset and scenario effects included in the modeling framework.
- Distributional analysis: ECDF analyses compare predicted P(Risky) for feature-present versus feature-absent queries using KS distance and ∆median.Length analyses use 30 equal-mass token-length bins to compare empirical Risky rates across feature states and scenarios.
- Propensity and uplift analysis: Propensity models estimate feature presence from other features and dataset/scenario indicators, with overlap diagnostics determining when uplift estimates are reported.IPW and propensity-stratified contrasts are treated as quasi-causal only under adequate overlap, and otherwise interpreted as associational.
- Propensity and uplift analysis: Feature-specific overlap-conditioned uplift results are summarized alongside ECDF separations in tables documenting common support and estimated effects.The uplift table reports percentage-point changes in Pr(RISKY) under IPW and propensity-stratified matching, omitting poor-overlap cases.
Hallucination LLM-Judge
The hallucination judge grades model answers against query-specific gold references using structured correctness decisions. Its examples cover exact matches, semantic equivalence, acceptable set members, contradictions, and unsupported or invented claims.
- Judge design: The judge returns a binary correctness decision and concise rationale for each query–answer–gold-reference example.The rationale should identify the decisive match or mismatch.
- Evaluation criteria: Semantic equivalence is accepted even when an answer is phrased differently from the gold reference.The rubric also accepts any member of an explicitly provided gold set.
- Evaluation criteria: Invented facts, contradictions, and unsupported specifics are penalized under the grading criteria.A nonexistent monarch and an incorrect numerical value illustrate contradiction-based rejection.
- Evaluation examples: A semantically equivalent photosynthesis explanation and a valid prime-number answer are labeled correct despite differing surface forms or allowable answers.These examples demonstrate the rubric’s handling of paraphrase and set-valued references.
- Diagnostic analyses: The paper reports heatmaps, reliability curves, propensity-overlap plots, probability KDEs, and ECDFs to characterize feature prevalence, calibration, support, class separation, and predicted risk.The heatmap compares extractive, multiple-choice, and abstractive queries across Safe, Borderline, and Risky categories.