XYBench: Can LLMs Respond Pragmatically to Queries with Misconceptions?

• The XY problem arises when users tunnel on a chosen method rather than their underlying goal; pragmatic help identifies procedural misconceptions and redirects toward that goal.This focus differs from work on false presuppositions limited to factoid beliefs and instruction-following that rewards literal compliance regardless of intent.Source: §1 · pp. 1–2
• XYBENCH contains 8,115 human-validated queries: 1,741 from Stack Overflow and Stack Exchange, 6,272 synthetic WikiHow queries, and 102 manually curated queries.The manually curated examples come from other online community forums.Source: §1 · p. 2
• Each query covers solutions to the stated request, identification of its misconception, and pragmatic alternatives addressing the underlying goal.These responses come from human answers for Stack Overflow, Stack Exchange, and the curated subset, and a multi-stage pipeline for WikiHow; all queries and metadata undergo filtering and human validation.Source: §1 · p. 2
• LLMs address S-requests in 0.75–0.92 but I-requests in 0.33–0.71, and identify misconceptions at most 60–63% versus humans’ 79–90%.On the unseen manually curated subset, models identify misconceptions at most 52% versus 79% for humans.Source: §1 · p. 2
• Models prefer pragmatic responses in multiple choice (up to 84%), while explicitly providing the S-request and I-request improves generation without closing the human-performance gap.The findings identify intent and misconception detection as bottlenecks, alongside difficulty leveraging them in response generation.Source: §1 · p. 2

2.1 Ambiguity and False Presuppositions

• Prior work studies ambiguous queries and false presuppositions, including LLMs’ tendency to endorse false claims.The cited lines cover multiple interpretations, underspecified or poorly posed questions, unanswerable questions, presupposition detection and correction, and sycophantic endorsement.Source: §2.1 · p. 2
• XYBENCH is presented as the first benchmark to target procedural misconceptions—problem framings about decomposition or method choice—that factual correctness checks miss.The authors argue this capability is increasingly important as LLMs are deployed as assistive agents.Source: §2.1 · p. 2

2.2 Misconceptions and Cooperative Systems

• Cooperative response theory distinguishes a posed question from the underlying need, calling for systems to identify misconceptions and address intended problems rather than reinforce false assumptions.A misconception is an internally coherent but incorrect or suboptimal problem framing that can bias the asker toward ineffective solutions.Source: §2.2 · p. 3
• Although LLMs serve as expert systems where misconception-driven queries are documented, they struggle with pragmatic, theory-of-mind, and non-literal intent, and are often trained to follow instructions literally.Prior work on latent misconceptions instead focuses on modeling student errors; XYBENCH examines cooperative responses when user misconceptions appear.Source: §2.2 · p. 3

3.1 Structure and Terminology

• Each instance pairs a misconceived stated request (S-request) with an intended request (I-request) that is not explicit but can be inferred from context, and evaluates three response dimensions.These are S-solutions, identification of the S–I misconception, and I-solutions that address the underlying goal.Source: §3.1 · p. 3

3.2 Stack Overflow & Stack Exchange

• The SO/SE pipeline mines XY-problem threads, extracts stated and intended requests using human answers as evidence, and filters out queries for which a literal answer is more helpful.It retains 1,557 SO and 184 SE queries; validation found 91–95% extraction accuracy against human judgments and 84.0% of instances fully correct across fields.Source: §3.2 · pp. 3–5

3.3 Manually-Curated

• The authors manually curated 102 in-the-wild queries, then verified each instance’s requests and response dimensions; no MCQ filter was needed because the misconception was present by construction.Queries were minimally edited, and only those where answering the stated request alone left the underlying problem unresolved were retained.Source: §3.3 · p. 5

3.4 WikiHow

• For WikiHow, an LLM generated queries from likely novice mistakes, refined inferable intended requests, added psychological distractors, and removed weak-gap cases, yielding 6,272 queries.Two psychologically grounded variations were generated per query; a 400-instance human validation found 68.5% fully correct across all fields.Source: §3.4 · pp. 5–6

3.5 Data Statistics

• 8,115 queries make up XYBENCH across Stack Overflow, Stack Exchange, WikiHow, and manually curated sources; SO and SE misconceptions most often involve approach fixation (42.8–59.2%).Other SO and SE misconception types are missing knowledge (25–31.5%), false belief (9.8–15.5%), and tool and syntax fixation (6.0–10.1%).Source: §3.5 · p. 6

3.6 Metrics for Generative Evaluation of Responses

• The metrics assess whether responses include stated and intended solutions, whether they emphasize the intended solution, and whether they identify the misconception explicitly or implicitly.Fully Pragmatic responses correct the misconception and focus on the underlying goal; Fully Cooperative responses also engage with the literal request.Source: §3.6 · pp. 6–7
• Human validation supports Claude Sonnet 4.5 as an evaluation judge, with inter-annotator exact agreement ≥0.955 and judge–human exact agreement ≥0.742 across dimensions.Results largely did not vary when GPT-4.1 or GPT-5-mini was used as judge.Source: §3.6 · p. 7

4 Results

• Tagging individual approach units to judge misconception identification yielded substantially better agreement with human annotations than judging full responses.The initial full-response judgment was unreliable because even strong models lacked sufficient pragmatic capability for the task.Source: §4 · p. 7

Models reliably address the stated request but rarely address the users’ underlying goal

• Models trail human answers in misconception identification by 25–40 points on SO, SE, and Manual, and identifying a misconception still does not ensure inclusion of the I-solution.When models identify the misconception, they include the I-solution in 61–85% of responses on SO, SE, and Manual, but in as few as 36% on WikiHow.Source: Models reliably address the stated… · p. 7
• Models include stated-request solutions in 0.57–0.91 of responses but include implicit-goal solutions in only 0.33–0.71, and prioritize them less often than humans.On SO and SE, pragmatic orientation is 22–41% for models versus 52–60% for humans; on the Manual set, it is 29–36% versus 82%.Source: Models reliably address the stated… · p. 7, Table 2 · p. 8
• Highest-scored human answers are marginally more fully pragmatic than accepted answers: 0.47 vs. 0.45 on SO and 0.56 vs. 0.54 on SE.The paper reports that highest-scored answers reflect what the broader community valued, while accepted answers are chosen by the question poster.Source: Models reliably address the stated… · p. 8
• In multiple-choice evaluation, models choose pragmatic responses in 71–84% (SO), 61–79% (SE), 64–79% (Manual), and 56–83% (WikiHow), despite struggling to generate them.The choices contrasted a primarily literal S-solution response with one focused on misconception identification and the I-solution.Source: Models reliably address the stated… · p. 8

Detecting intent is not the only bottleneck

• With S-request and I-request supplied at generation time, I-solution emphasis improves up to twice as often, yet only Gemini 3.1 Pro closes the human gap on SO and SE.Models still lag human references in misconception identification; the findings distinguish detecting intent from using it to steer generation when it diverges from the stated request.Source: Figure 3 · p. 9, Detecting intent is not the only… · p. 9

5 Conclusion & Discussion

• XYBench comprises 8,115 queries whose stated requests reflect misconceptions about users’ underlying goals, drawn from technical and everyday domains.The benchmark combines real Stack Overflow and Stack Exchange conversations, synthetically constructed WikiHow queries, and a manually curated everyday subset.Source: §5 · p. 9
• Explicit S- and I-request provision lowers misconception identification for Claude Sonnet 4.6 and GPT-5 on SO.The authors suggest SO may be heavily memorized by these models, leading them to treat queries as resolved dual-answer tasks rather than reason through divergent goals.Source: §5 · p. 9
• Models usually answer literal requests, rarely identify misconceptions, and trail humans in pragmatic redirection, despite preferring pragmatic choices and improving with explicit intent or correction instructions.In multiple-choice evaluation, models predominantly choose the pragmatic response; generation improves with explicit intent or correction instructions but still lags behind human responses.Source: §5 · p. 9
• The authors argue that pragmatic capability requires intent modeling, while instruction-following and preference data reinforce literal compliance.Identifying misconceptions also requires better theory of mind and social-world modeling (Zhou et al., 2025); the tendency relates to sycophancy and utility optimization (Shapira et al., 2026; Sun and Wang, 2026).Source: §5 · p. 9

6 Limitations

• The benchmark lacks human judgments of naturally occurring misconception queries and human references for synthetic WikiHow queries, leaving expertise- and culture-based response differences for future study.Source: §6 · p. 10
• Its platform sources may overrepresent procedure-focused problems, limiting exhaustive coverage of real-world query types despite naturalistic grounding across technical and everyday domains.Source: §6 · p. 10
• WikiHow queries are synthetically generated and may differ distributionally from natural XY problems; literature-grounded generation and manual curation mitigate this, while mining naturally elicited queries remains future work.Source: §6 · p. 10
• English-only evaluation limits generalizability because misconception-response pragmatics may vary across languages and cultures; the authors call for cross-lingual and cross-cultural study.Source: §6 · p. 10

Manually-Curated Queries and References

• The manually curated queries may reflect the authors’ intuitions and only a subset of misconceptions across domains and expertise levels; automatic in-the-wild identification is a future direction.Source: Manually-Curated Queries and References · p. 10

A Example Query, Human and Model Responses

• Figures 4–6 present Stack Overflow, Stack Exchange, and manually curated query examples with model responses and human references, while Figure 7 shows a WikiHow example with model responses.Source: A Example Query, Human and Model… · p. 14

B.1 Prompts used for extraction of query and response metadata, and for MCQ filtering

• The extraction prompts derive rewritten queries, S-requests, and I-requests from original queries and human answers, then assess S-solutions, I-solutions, and misconception identification.A separate oracle MCQ prompt filters out queries better aligned with an S-solution and not needing an I-solution.Source: §B.1 · p. 14

B.2 Human Validation MCQ Filtering of Stack Overflow and Stack Exchange

• On 100 balanced filtering decisions, two CS-background authors had substantial agreement (κ = 0.764), and GPT-5 matched their judgments 93.0% on average.The authors independently labeled each decision agree, borderline, or disagree; Table 3 reports pipeline accuracy.Source: §B.2 · p. 14, Table 3 · p. 14

B.3 Human Validation of extracted query and response dimension of Stack Overflow

• Across five extracted fields, strict accuracy was 91.0–95.0% per dimension, and 84.0% of sampled instances were correct on all five simultaneously.The dimensions were S-request, I-request, S-solution, I-solution, and S-I-misconception-identification.Source: §B.3 · p. 14
• For 100 sampled instances, annotator exact agreement was 0.970–1.000 and Krippendorff’s α was 0.687–1.000 across the five extraction dimensions.Two authors independently scored each dimension as correct, partial, or incorrect.Source: §B.3 · p. 14

B.4 Single-turn Capability

• Only 31.8% of SO/SE threads had asker follow-ups; among 50 sampled follow-ups, 42% accepted the reframe, 26% negotiated without resolution, and 32% were unrelated.The authors interpret these results as evidence that misconceptions are most often resolved in the initial expert response, supporting the single-turn design.Source: §B.4 · p. 14

C.1 Prompts used WikiHow Guided Data Generation and Automated Filtering and Rewriting

• The initial WikiHow generation prompt specifies a query, surface and intended requests and solutions, and misconception identification.These components are shown in Figure 11.Source: §C.1 · p. 14
• Generated queries are checked for expert identifiability and obvious clues, with rewriting or clue removal repeated up to three times when needed.The I-request is rewritten if a domain expert cannot identify it; queries with obvious clues are revised to remove them.Source: §C.1 · p. 14
• A further filtering and regeneration step removes obvious clues and rewrites queries with a psychologically grounded explanation motivating the surface request.The passage describes this process as applying to given queries, followed by rewriting.Source: §C.1 · pp. 14–15
• The query filtering pipeline also assesses the quality of the gap between surface and intended requests.Figures 15’s prompt is described as filtering queries based on S- and I-request gap quality.Source: §C.1 · p. 15

C.2 Human Validation Setup

• Two authors independently annotate 400 generated instances across six dimensions, scoring each as correct, partial, or incorrect.Dimensions cover correct capture of the S-request and I-request, query plausibility, and correctness of the S-solution, I-solution, and misconception identification.Source: §C.2 · p. 16
• Inter-annotator agreement is reported using pairwise exact agreement and Krippendorff’s α, while pipeline accuracy is reported under strict and lenient thresholds.Strict counts only correct responses; lenient also counts partial responses as correct.Source: §C.2 · pp. 16–17

D Data Distribution of Stack Overflow and Stack Exchange

• The taxonomy was developed by manually coding 100 sampled LLM-labeled queries, then using the resulting four categories to re-tag queries.The categories describe Approach Fixation, Missing Knowledge, False Belief, and Tool & Syntax Fixation.Source: D Data Distribution of Stack Overflow… · p. 17
• For real Stack Overflow and Stack Exchange threads, a prompt extracts S-solution, I-solution, and misconception identification given the query and S- and I-requests.Source: Figure 9 · p. 18
• A filtering prompt presents shuffled S-solution and I-solution choices to select queries better suited to an S-solution.The prompt asks models to choose the response option most helpful to the user.Source: Figure 10 · p. 19
• The WikiHow generation prompt produces an initial query, S- and I-requests and solutions, and an S-I misconception identification.Source: Figure 11 · p. 19
• WikiHow queries are checked for identifiable I-requests and obvious clues, then filtered or rewritten to add psychologically grounded motivation for the S-request.The passages describe expert review of I-request identifiability, clue removal, and filtering and regeneration.Source: Figure 12 · p. 20, Figure 13 · p. 20, D Data Distribution of Stack Overflow… · p. 21
• Human validation covers 400 sampled WikiHow instances across six dimensions, with strict accuracy counting only correct or yes and lenient accuracy also counting partial responses.Source: Table 5 · p. 22
• Approach Fixation is the dominant misconception category in both forums, reaching 59.2% in Stack Exchange.The distribution is reported across Stack Overflow and Stack Exchange; Missing Knowledge ranks second, while Tool & Syntax Fixation and False Belief make up a smaller but meaningful share.Source: D Data Distribution of Stack Overflow… · p. 17, Table 6 · p. 22

E.1 Reasoning for Decomposed Metrics for Generative Evaluation of LLM Responses

• Human reference answers ground extraction of the S-request, I-request, S-solution, I-solution, and S-I-misconception-identification dimensions.The task extracts these dimensions from accepted answers and the top three highest-scored answers; human validation found the procedure highly reliable across all five dimensions.Source: §E.1 · p. 22
• Verbose LLM responses, alternate suggestions, hedging, and claims about the “best” approach make similar evaluation challenging because even strong LLM judges are easily distracted by surface signals.Source: §E.1 · p. 22
• The evaluation decomposes responses into what was recommended (Presence), how prominently it was recommended (Emphasis), and whether the misconception was flagged (Misconception Identification).This setup measures the three dimensions independently of an LLM judge’s pragmatic capabilities and uses local classification of short extracted units.Source: §E.1 · p. 22
• Human validation across benchmark construction and LLM-as-judge setups confirms that pipeline quality is not contingent on the choice of Claude Sonnet 4.5.The model was used for its strong instruction-following in data extraction and evaluation.Source: §E.1 · p. 22

E.2 Human Validation Setup for Claude Sonnet 4.5

• The human-validation sample comprised 500 responses—100 Stack Overflow and 400 WikiHow—independently annotated by two authors and judged with Claude Sonnet 4.5.The study compared inter-annotator and judge–human agreement for each evaluation dimension and dataset.Source: §E.2 · p. 22
• Human annotators agreed perfectly on solution presence and emphasis (pairwise agreement = 1.00, Krippendorff’s α = 1.00), while misconception identification agreement was strong but lower.For S-I-misconception-identification, pairwise agreement was 0.955–0.991 and Krippendorff’s α was 0.910–0.979, reflecting subjectivity in judging whether a misconception was adequately addressed.Source: §E.2 · p. 22
• Claude Sonnet 4.5’s agreement with humans was highest for presence detection (pairwise agreement ≥0.975, Krippendorff’s α ≥0.915) and lower for emphasis and misconception identification.Emphasis agreement was 0.742–0.798 pairwise and 0.587–0.600 by Krippendorff’s α; misconception-identification agreement was 0.846–0.933 and 0.634–0.866, respectively. Overall, human–model agreement was commensurate with human–human agreement, supporting Claude Sonnet 4.5 as a judge proxy.Source: §E.2 · pp. 22–23
• On 250 validation pairs, gpt-4.1 and gpt-5-mini showed moderate-to-almost-perfect or fair-to-almost-perfect agreement with Claude Sonnet 4.5 and human annotations, respectively.gpt-4.1 had κ = 0.45−0.76 agreement with Claude and κ = 0.44−0.87 with humans; gpt-5-mini had κ = 0.31−0.73 with Claude and κ = 0.38−0.86 with humans. They tended to overflag different evaluation dimensions, and Claude agreed best with human annotations.Source: §E.2 · p. 23

F.1 Relationship between Misconception identification and mention of I-solution

• The relationship analysis recorded binary signals for misconception identification and I-solution inclusion for each model and query, then compared conditional I-solution probabilities.The analysis assessed the association’s significance and reported results in Table 10.Source: §F.1 · p. 23
• Models were more likely to mention the I-solution when they identified the misconception, but recognition did not guarantee that they included it.The analysis compared conditional I-solution presence for responses with and without misconception identification.Source: §F.1 · p. 23

F.2 Multiple Choice Evaluation Setup

• The multiple-choice setup pairs each query with an S-solution response and a response combining misconception identification with I-solution, in randomized order.Claude-Sonnet-4.5 generated the responses from explicit response dimensions; the evaluated model rewrote them to similar lengths without adding information.Source: §F.2 · p. 23
• In multiple-choice evaluation, models overwhelmingly prefer the pragmatic response on Stack Exchange and Stack Overflow, and largely prefer it on the manually curated set and WikiHow.The four figures report results for Stack Exchange, Stack Overflow, the manually curated set, and WikiHow, respectively.Source: Figure 17 · p. 25, Figure 18 · p. 25, Figure 19 · p. 26

G License, Intended Use, Information About Use Of Ai Assistants

• XYBench will be released under CC BY 4.0; its public-source data uses permissible licenses, and the anonymized manually curated subset contains no personally identifiable information.Claude Code assisted with collaborative processing of Stack Overflow data.Source: G License, Intended Use, Information… · p. 27

H Human Annotation Interface

• The human annotation interface covers request correctness, response-dimension correctness, solution-presence judgments, emphasis orientation, and misconception identification.Figures 21–25 show the annotation framework.Source: H Human Annotation Interface · p. 27, Figure 21 · p. 30, Figure 22 · p. 30

F.3 Explicitly providing S-request and I-request at generation time substantially improves performance

• With explicit requests, models include the I-solution roughly twice as often on WikiHow and answer the intended request nearly four times as often, but still trail humans.The oracle experiment supplied both S-request and I-request alongside the query.Source: §F.3 · p. 23, Table 11 · p. 28
• On SE, oracle prompting raises misconception identification by 33% and I-solution presence by 38% on average; on SO, I-request emphasis rises 25% while identification falls 28–45% for all but Gemini 3.1 Pro.The authors suggest SO models may treat the explicit requests as a resolved dual-answer task rather than reason through why the goals diverge.Source: §F.3 · pp. 23–24, Table 11 · p. 28
• The oracle effect is weaker on the manually curated set, where I-solution gains are modest and S-solution is near ceiling; models still emphasize S-solution more than humans do.The results point to two bottlenecks: detecting communicative intent and using it to steer generation when it diverges from the stated request.Source: §F.3 · p. 24, Table 11 · p. 28

F.4 Adding “please correct the request if there are any misconceptions” along with the query improves performance but does not close the gap

• The misconception-correction prompt raises identification by 6–26 points across datasets, but I-solution presence improves only 2–8 points and pragmatic emphasis changes minimally.Models still lag humans by up to 13 points on SO and 22 points on SE in misconception identification.Source: §F.4 · p. 27, Table 12 · p. 29