The DOI matches its Crossref record; 29 of the 36 references it cites were found in a registry or an index; 9 of the 20 sources its claims cite were checked for support.
Flaw or Artifact? Rethinking Prompt Sensitivity in Evaluating LLMs
Open accessAbstract
Prompt sensitivity, referring to the phenomenon where paraphrasing (i.e., repeating something written or spoken using different words) leads to significant changes in large language model (LLM) performance, has been widely accepted as a core limitation of LLMs.In this work, we revisit this issue and ask: Is the widely reported high prompt sensitivity truly an inherent weakness of LLMs, or is it largely an artifact of evaluation processes?To answer this question, we systematically evaluate 7 LLMs (e.g., GPT and Gemini family) across 6 benchmarks, including both multiple-choice and open-ended tasks on 12 diverse prompt templates.We find that much of the prompt sensitivity stems from heuristic evaluation methods, including log-likelihood scoring and rigid answer matching, which often overlook semantically correct responses expressed through alternative phrasings, such as synonyms or paraphrases.When we adopt LLM-as-a-Judge evaluations, we observe a substantial reduction in performance variance and a consistently higher correlation in model rankings across prompts.Our findings suggest that modern LLMs are more robust to prompt templates than previously believed, and that prompt sensitivity may be more an artifact of evaluation than a flaw in the models.
Overview
AI snapshot
Written from the paper’s full text.
TL;DR
The paper argues that reported prompt sensitivity in LLM evaluation is largely an artifact of heuristic scoring, not an inherent model flaw. Using LLM-as-a-Judge across 7 LLMs, 6 benchmarks, and 12 prompt templates, it finds much lower performance variance and much higher ranking consistency than heuristic evaluation.
Key findings
- Heuristic evaluation exaggerates prompt sensitivity. On ARC-Challenge, Gemma-2.0’s accuracy ranges from 0.25 to 0.90 with standard deviation 0.28 under heuristics, versus a 0.17 range and standard deviation 0.005 under LLM-as-a-Judge.
- Ranking consistency improves sharply under LLM-as-a-Judge. On ARC-Challenge, mean Spearman rank correlation among open-source models rises from 0.30/0.31 under heuristics to 0.92 with LLM-as-a-Judge, and to 0.95 when proprietary models are included. Table 1 reports ARC-Challenge ρ from 0.3036 to 0.9546, OpenbookQA from 0.4212 to 0.9386, GPQA Diamond from 0.1542 to 0.8960, NarrativeQA from 0.5927 to 0.8662, and MATH from 0.9593 to 0.9647.
- Human annotations support the LLM-judge results. Across six datasets, human agreement has Fleiss’ κ over 0.6, combined agreement 0.8132, and combined perfect agreement 73%; human-majority vs LLM-judge combined Cohen’s κ is 0.9247.
- The pattern also appears in older models. For Llama-2 and Mistral, LLM-as-a-Judge reduces standard deviation across paraphrased prompts from 0.1413 to 0.0182 and from 0.0389 to 0.0098, respectively.
Methods
- Evaluated 7 LLMs: LLaMA-3.1-8B-Instruct, Qwen2-7B-Instruct, Gemma-2-9B-it, Ministral-8B-Instruct-v0.2, GPT-4o-mini, GPT-4.1-mini, and Gemini 2.0 Flash. Benchmarks: ARC-Challenge, GPQA-Diamond, OpenbookQA, NarrativeQA, MATH, and SimpleQA.
- Constructed 12 prompt templates per benchmark using GPT-4o paraphrasing; multiple-choice datasets shared a pool of 12 templates, while open-ended tasks used 12 benchmark-specific templates. Compared heuristic evaluation (log-likelihood for multiple-choice; format-specific extraction/normalization for NarrativeQA and MATH; no parser for SimpleQA) with LLM-as-a-Judge, defaulting to Gemini 2.0 Flash and checking GPT-4o-mini on ARC-Challenge.
- Measured prompt sensitivity via standard deviation of accuracy across prompt templates and mean pairwise Spearman rank correlation across prompts. Human validation used three undergraduate annotators, 50 sampled questions per dataset, one model under all 12 prompts, yielding 600 answers per dataset-model pair and 10,800 annotations total, with majority voting.
Limitations
- The paper states computational constraints limited evaluation to 12 prompt templates per benchmark, though it reports that ARC-Challenge results with 12 prompts closely match those with over 100 prompts.
- Evident from the text: SimpleQA has no official heuristic parser, so it is evaluated only with LLM-as-a-Judge; NarrativeQA lacks results for Mistral and Qwen due to long-context requirements; and the human study uses three undergraduate annotators from UCSB, which may limit annotator diversity.
Model-written summary. Check it against the paper before relying on it.
Integrity signals
Registry
Cites
Statistics
Verification
Claims the paper makes
9 of 20 cited sources checked · 5 supported · 0 flagged · accuracy not yet measured
- Premisecitation
Previous studies typically rely on heuristic evaluation, such as regular-expression-based answer extraction or log-likelihood scoring over candidates.
“To investigate this, we find that previous studies (Voronov et al., 2024;Chatterjee et al., 2024) typically rely on heuristic evaluation, such as regular-expression-based answer extraction or loglikelihood scoring over candidates.”
Quote in the paperSubstantial claim2 of 2 sources checked- cites Voronov et al., 2024 source
Section 4.2 discusses DIRECT, CHANNEL, and CALIBRATION prediction methods, while Section 5.5 discusses sensitivity to prompt templates. Neither passage establishes that previous studies typically use heuristic evaluation or mentions regular-expression answer extraction or log-likelihood scoring over candidates. Searches for those terms and related descriptions found no passage addressing the claim. (LLM, conf=0.72)
Experimental — accuracy not yet measured - cites Chatterjee et al., 2024 source
The Introduction says earlier prompt-sensitivity studies used measures such as accuracy, exact match, performance spread, and variation ratio. It does not describe regular-expression-based answer extraction or log-likelihood scoring over candidates; its discussion of log-likelihood concerns responses to prompt variations, not candidate scoring. (LLM, conf=0.84)
Experimental — accuracy not yet measured
- Premisecitation
The relative rankings of LLMs can shift substantially depending on the prompt template used.
“More critically, the relative rankings of LLMs can shift substantially depending on the prompt template used (Polo et al., 2024;Mizrahi et al., 2024).”
Quote in the paperSubstantial claim1 of 2 sources checked - Findingown data
On NarrativeQA, the correlation rises from 0.40 (heuristics) to 0.87 (LLM-as-a-Judge).
“On NarrativeQA, the correlation rises from 0.40 (heuristics) to 0.87 (LLM-as-a-Judge).”
Quote in the paperSubstantial claim - Findingown data
Heuristic evaluation methods often exaggerate the prompt sensitivity of LLMs.
“We discover that heuristic evaluation methods often exaggerate the prompt sensitivity of LLMs.”
Quote in the paperSubstantial claim - Findingown data
The heuristic methods exhibit significantly greater sensitivity to prompt variation than LLM-as-a-Judge.
“When comparing the performance of the model under heuristic evaluation and LLM-as-a-Judge, we find that the heuristic methods exhibit significantly greater sensitivity to prompt variation (Figure 2).”
Quote in the paperSubstantial claim
- Findingown data
The majority-voted outcomes closely align with the LLM-judge results.
“Moreover, the majority-voted outcomes closely align with the LLM-judge results, underscoring the reliability of LLM-as-a-Judge evaluation.”
Quote in the paperSubstantial claim - Findingown data
Human annotators' judgments are highly consistent.
“Their judgments are highly consistent, and majority voting is used to resolve the few cases of disagreement.”
Quote in the paperSubstantial claim - Findingnone given
The standard deviation of the LLM-as-a-Judge method is always low.
“The standard deviation of the LLM-as-a-Judge method is always low.”
Quote in the paperNot a substantial claim - Findingown data
On ARC-Challenge, the performance of Gemma-2.0 varies widely across prompts, with accuracy ranging from 0.25 to 0.90 and a high standard deviation of 0.28.
“For instance, on ARC-Challenge, the performance of Gemma-2.0 varies widely across prompts, with accuracy ranging from 0.25 to 0.90 and a high standard deviation of 0.28.”
Quote in the paperSubstantial claim - Findingown data
Well-designed heuristic methods show low prompt sensitivity similar to LLM-as-a-Judge.
“Well-designed heuristic methods show low prompt sensitivity similar to LLM-as-a-Judge.”
Quote in the paperSubstantial claim - Findingown data
The Spearman rank correlation measuring model performance rankings across the four open-source models improves from 0.31 under heuristic evaluation to 0.92 with LLM-as-a-Judge.
“Furthermore, the Spearman rank correlation measuring model performance rankings across the four open-source models remarkedly improves from 0.31 under heuristic evaluation to 0.92 with LLM-as-a-Judge.”
Quote in the paperSubstantial claim - Findingown data
Results are stable across scales.
“However, we find that results are stable across scales.”
Quote in the paperSubstantial claim - Findingown data
Newly proposed benchmarks exhibit low prompt sensitivity.
“Newly proposed benchmarks exhibit low prompt sensitivity.”
Quote in the paperSubstantial claim - Findingown data
The rate of perfect agreement is notably high.
“The rate of perfect agreement is also notably high (Table 3).”
Quote in the paperSubstantial claim - Findingown data
Mean accuracy improves under LLM-as-a-Judge.
“Additionally, mean accuracy improves under LLM-as-a-Judge, suggesting heuristic methods often miss valid answers due to overly rigid extraction rules.”
Quote in the paperSubstantial claim - Findingown data
Using LLM-as-a-Judge significantly reduces variance across paraphrased prompts for older instruction-tuned LLMs (Llama-2-7B-Chat and Mistral-7B-Instruct-v0.1).
“From Table 5, we find that using LLM-as-a-Judge significantly reduces variance across paraphrased prompts.”
Quote in the paperSubstantial claim - Premisecitation
Simply changing the option format from letters to numbers completely reverses the ranking order of four evaluated open-source models in ARC-Challenge.
“For example, simply changing the option format from letters (e.g.,”
Quote in the paperSubstantial claim0 of 1 sources checked- cites Clark et al., 2018 source
- Interpretationreasoning
Heuristic methods often miss valid answers due to overly rigid extraction rules.
“Additionally, mean accuracy improves under LLM-as-a-Judge, suggesting heuristic methods often miss valid answers due to overly rigid extraction rules.”
Quote in the paperSubstantial claim - Interpretationown data
Modern LLMs exhibit less prompt-induced variance than previously reported.
“This further supports the conclusion that modern LLMs exhibit less prompt-induced variance than previously reported.”
Quote in the paperSubstantial claim- Rests on an assumption: The seven evaluated LLMs are representative of modern LLMs.
- Rests on an assumption: The six benchmarks are representative of the tasks used to evaluate LLMs.
- Interpretationown data
LLM-as-a-Judge evaluation is reliable.
“Moreover, the majority-voted outcomes closely align with the LLM-judge results, underscoring the reliability of LLM-as-a-Judge evaluation.”
Quote in the paperSubstantial claim- Rests on an assumption: Human annotations provide a reliable ground truth for answer correctness.
- Interpretationown data
With sufficient domain-specific prompt engineering, heuristic methods can provide stable evaluations.
“These results indicate that, with sufficient domain-specific prompt engineering, heuristic methods can provide stable evaluations.”
Quote in the paperSubstantial claim - Interpretationown data
Prompt templates have minimal impact on the answer correctness.
“This further reinforces our conclusion that prompt templates have minimal impact on the answer correctness.”
Quote in the paperSubstantial claim - Interpretationreasoning
Heuristic evaluation approaches may introduce errors when model outputs deviate from expected formats.
“These heuristic evaluation approaches, though historically popular due to their simplicity (Zellers et al., 2019;Reddy et al., 2019;Brown et al., 2020;Hendrycks et al., 2021a), may introduce errors when model outputs deviate from expected formats.”
Quote in the paperSubstantial claim - Interpretationown data
Model performance remains stable across different prompt variations.
“These results indicate that model performance remains stable across different prompt variations.”
Quote in the paperSubstantial claim - Study design and conductown data
Human annotations (after majority voting) exhibit strong agreement with the LLM-judge (Table 4).
“Furthermore, human annotations (after majority voting) exhibit strong agreement with the LLM-judge (Table 4).”
Quote in the paperSubstantial claim - Interpretationown data
The high consistency between LLM-judges and human annotators strongly indicates that prompt sensitivity is largely an artifact of the evaluation method rather than an inherent flaw in LLMs.
“The high consistency between LLM-judges and human annotators strongly indicates that prompt sensitivity is largely an artifact of the evaluation method rather than an inherent flaw in LLMs.”
Quote in the paperSubstantial claim- Rests on an assumption: Human annotations provide a reliable ground truth for answer correctness.
- Study design and conductown data
Inter-annotator agreement is high (Table 4).
“Inter-annotator agreement is high (Table 4), indicating the overall quality and consistency of human annotations.”
Quote in the paperSubstantial claim - Interpretationown data
Prompt sensitivity observed in prior work is largely an artifact of heuristic evaluation, not an inherent flaw of LLMs.
“These findings suggest that prompt sensitivity observed in prior work is largely an artifact of heuristic evaluation, not an inherent flaw of LLMs.”
Quote in the paperSubstantial claim - Interpretationown data
The LLM-judge provides reliable correctness judgments when comparing model answers against ground truth across diverse prompt templates.
“This demonstrates that the LLM-judge provides reliable correctness judgments when comparing model answers against ground truth across diverse prompt templates.”
Quote in the paperSubstantial claim - Interpretationown data
The high inter-annotator agreement indicates overall quality and consistency of human annotations.
“Inter-annotator agreement is high (Table 4), indicating the overall quality and consistency of human annotations.”
Quote in the paperSubstantial claim - Interpretationreasoning
Models may generate correct answers, but because their outputs are not aligned with the rigid evaluation format, they are mistakenly marked as incorrect.
“More specifically, models may generate correct answers, but because their outputs are not aligned with the rigid evaluation format, they are mistakenly marked as incorrect.”
Quote in the paperSubstantial claim - Interpretationown data
Much of the observed prompt sensitivity in LLM evaluations is not due to inherent model weaknesses, but rather an artifact introduced by heuristic evaluation methods.
“In this work, we demonstrate that much of the observed prompt sensitivity in LLM evaluations is not due to inherent model weaknesses, but rather an artifact introduced by heuristic evaluation methods.”
Quote in the paperSubstantial claim- Rests on an assumption: The seven evaluated LLMs are representative of modern LLMs.
- Rests on an assumption: The six benchmarks are representative of the tasks used to evaluate LLMs.
- Interpretationown data
The reduction of prompt-induced variability is not due to recent models 'fixing' the issue.
“These additional results indicate that the reduction of prompt-induced variability is not due to recent models”
Quote in the paperSubstantial claim - Interpretationown data
Answer correctness does not vary substantially across different prompt templates.
“This provides strong evidence that the answer correctness does not vary substantially across different prompt templates.”
Quote in the paperSubstantial claim - Premisereasoning
Instruction-tuned LLMs are explicitly optimized to handle a wide range of input formats.
“this remains counterintuitive given that instructiontuned LLMs are explicitly optimized to handle a wide range of input formats.”
Quote in the paperSubstantial claim
- Probably not a substantial claim. 14
- It may state two separate claims as one. 12
- The quote only partly states this claim. 1
Probably not a substantial claim.
PremisecitationThe ranking inconsistency with diverse prompt templates has been widely reported.
“The ranking inconsistency with diverse prompt templates has been widely reported (Polo et al., 2024;Mizrahi et al., 2024;Chatterjee et al., 2024;Sclar et al., 2023).”
Quote in the paperSubstantial claim2 of 3 sources checked- cites Polo et al., 2024 source
- cites Mizrahi et al., 2024
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
- cites Chatterjee et al., 2024 source
The Introduction (page unavailable) reports that performance varies across prompt templates, and Section 5.5 finds template variation especially consequential for multiple-choice tasks. But the paper does not report that the relative ranking of models changes across templates, which is the specific inconsistency claimed. (LLM, conf=0.83)
Experimental — accuracy not yet measured
Probably not a substantial claim. The paper’s conclusions may not depend on it.
PremisecitationVoronov et al. (2024) showed that no prompt format consistently performs best across models.
“Voronov et al. (2024) further showed that no prompt format consistently performs best across models.”
Quote in the paperNot a substantial claim1 of 1 sources checked- cites Voronov et al., 2024 source
It may state two separate claims as one.
PremisecitationLLM judges can better handle various output formats, paraphrasing, and ambiguous cases, making them more aligned with human evaluation.
“Compared to heuristics, LLM judges can better handle various output formats, paraphrasing, and ambiguous cases, making them more aligned with human evaluation (Liu et al., 2023;Kocmi and Federmann, 2023;Zheng et al., 2023;Wang et al., 2023).”
Quote in the paperSubstantial claim1 of 3 sources checked- cites Liu et al., 2023 source
The Introduction and Conclusion report that G-EVAL correlates more strongly with human judgments than baseline NLG evaluators, particularly on open-ended tasks. The paper does not establish that LLM judges achieve this by better handling various output formats, paraphrasing, or ambiguous cases; those specific capabilities are not demonstrated in the cited passages or the additional passages searched. (LLM, conf=0.86)
Experimental — accuracy not yet measured - cites Kocmi and Federmann, 2023
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
- cites Zheng et al., 2023
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
Probably not a substantial claim.
PremisecitationExisting studies have reported that LLMs are highly sensitive to prompt phrasing.
“Although existing studies have reported that LLMs are highly sensitive to prompt phrasing (Voronov et al., 2024;Mizrahi et al., 2024)”
Quote in the paperSubstantial claim1 of 2 sources checked- cites Voronov et al., 2024 source
- cites Mizrahi et al., 2024
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
Probably not a substantial claim. The paper’s conclusions may not depend on it.
PremisecitationPolo et al. (2024) estimates the distribution of accuracy across prompts to improve evaluation efficiency.
“For example, Polo et al. (2024) estimates the distribution of accuracy across prompts to improve evaluation efficiency.”
Quote in the paperNot a substantial claim1 of 1 sources checked- cites Polo et al., 2024 source
It may state two separate claims as one.
Findingown dataApplying prompt sensitivity analysis to SimpleQA, we observe a low standard deviation and a high Spearman rank correlation of 0.8121 across 12 prompt templates.
“Applying our prompt sensitivity analysis, we observe a low standard deviation and a high Spearman rank correlation of 0.8121 across 12 prompt templates.”
Quote in the paperSubstantial claimIt may state two separate claims as one.
Findingown dataUnder these conditions, prompt sensitivity results are comparable to those obtained via LLM-as-a-Judge evaluation, with similarly low accuracy variance and high ranking consistency.
“Under these conditions, we observe prompt sensitivity results that are comparable to those obtained via LLM-as-a-Judge evaluation, with similarly low accuracy variance and high ranking consistency.”
Quote in the paperSubstantial claimIt may state two separate claims as one.
Findingown dataBoth the standard deviation across different prompts and the ranking correlation remain consistent regardless of whether GPT or Gemini is used as the judge.
“As shown in Table 2, both the standard deviation across different prompts and the ranking correlation remain consistent regardless of whether GPT or Gemini is used as the judge.”
Quote in the paperSubstantial claimIt may state two separate claims as one.
Findingown dataHuman-annotated results show consistently high agreement on the answer correctness, with a high Fleiss' κ over 0.6.
“Observation 1 Human-annotated results show consistently high agreement on the answer correctness (Table 3), with a high Fleiss' κ over 0.6 (Fleiss, 1971).”
Quote in the paperSubstantial claim0 of 1 sources checked- cites Fleiss, 1971
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
It may state two separate claims as one.
Findingown dataWhen evaluated with LLM-as-a-Judge, models show far more consistent performance and stable rankings across prompts.
“When evaluated with LLM-as-a-Judge, models show far more consistent performance and stable rankings across prompts.”
Quote in the paperSubstantial claimIt may state two separate claims as one.
Findingown dataWhen evaluated using LLM-as-a-Judge, Gemma-2.0's accuracy varies by only 0.17 across the same set of prompt templates, with a much lower standard deviation of just 0.005.
“In contrast, when evaluated using LLM-as-a-Judge, its accuracy varies by only 0.17 across the same set of prompt templates, with a much lower standard deviation of just 0.005.”
Quote in the paperSubstantial claimIt may state two separate claims as one.
Findingown dataHuman-annotated answer correctness shows high agreement across prompt templates, with Fleiss' κ values ranging from 0.6708 to 0.8132 and perfect agreement rates from 52% to 88% across datasets.
“Dataset Agreement Perfect Agreement Arc-Challenge 0.7654 86% OpenbookQA 0.7250 80% GPQA-Diamond 0.6708 52% NarrativeQA 0.7258 66% MATH 0.7859 68% SimpleQA 0.7881 88% Combined 0.8132 73% Table 3: Correctness of answers shows minimal variation across prompt templates, based on human annotations.”
Quote in the paperSubstantial claimIt may state two separate claims as one.
Findingown dataOn ARC-Challenge, the average Spearman rank correlation across prompts among open-source models increases from 0.30 (heuristics) to 0.92 (LLM-as-a-Judge), and further to 0.95 when proprietary models are included.
“On ARC-Challenge, the average Spearman rank correlation across prompts among open-source models increases from 0.30 (heuristics) to 0.92 (LLM-as-a-Judge), and further to 0.95 when proprietary models are included.”
Quote in the paperSubstantial claimIt may state two separate claims as one.
Findingown dataOn ARC-Challenge, the ranking consistency and variance metrics using 12 prompts closely match those obtained with over 100 prompts, suggesting that our analysis is representative.
“For example, on ARC-Challenge, the ranking consistency and variance metrics using 12 prompts closely match those obtained with over 100 prompts, suggesting that our analysis is representative.”
Quote in the paperSubstantial claimIt may state two separate claims as one.
Findingnone givenA key factor in prompt sensitivity is the heuristic evaluation protocol itself, which often leads to misclassification of correct outputs and overstates prompt sensitivity.
“In contrast, we show that a key factor is the heuristic evaluation protocol itself, which often leads to misclassification of correct outputs and overstates prompt sensitivity.”
Quote in the paperNot a substantial claimIt may state two separate claims as one.
Findingown dataModel performance and rankings are substantially more stable and reliable than previously reported.
“Through comprehensive experiments using LLMas-a-Judge across multiple benchmarks and prompt templates, we reveal that model performance and rankings are substantially more stable and reliable than previously reported.”
Quote in the paperSubstantial claim- Rests on an assumption: The seven evaluated LLMs are representative of modern LLMs.
- Rests on an assumption: The six benchmarks are representative of the tasks used to evaluate LLMs.
Probably not a substantial claim.
PremisecitationInstruction-tuning datasets such as FLAN and Super-NaturalInstructions include a diverse collection of tasks with varying natural language prompt templates.
“For example, instruction-tuning datasets such as FLAN (Longpre et al., 2023) and Super-NaturalInstructions (Wang et al., 2022) include a diverse collection of tasks (e.g., question answering, summarization, classification) with varying natural language prompt templates (Zhang et al., 2023).”
Quote in the paperSubstantial claim0 of 3 sources checked- cites Longpre et al., 2023
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
- cites Wang et al., 2022
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
- cites Zhang et al., 2023
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
The quote only partly states this claim.
Interpretationown dataPrompt-induced variability has consistently been an artifact of evaluation methods rather than a fundamental inconsistency in model behavior.
“These additional results indicate that the reduction of prompt-induced variability is not due to recent models”
Quote in the paperSubstantial claimProbably not a substantial claim. The paper may give no reason why it holds.
Interpretationnone givenHeuristic evaluation methods often fail when model outputs deviate from expected formats.
“Heuristic evaluation methods often fail when model outputs deviate from expected formats.”
Quote in the paperNot a substantial claimProbably not a substantial claim. The paper’s conclusions may not depend on it.
Study design and conductown dataThe human annotators had no prior exposure to the study.
“We recruit three undergraduate students from UCSB with no prior exposure to this study.”
Quote in the paperNot a substantial claimProbably not a substantial claim. The paper’s conclusions may not depend on it.
Study design and conductown dataThe annotators were fluent in English.
“The undergraduate students are Asian or Asian American, and all of them are fluent in English.”
Quote in the paperNot a substantial claimProbably not a substantial claim. The paper’s conclusions may not depend on it.
Study design and conductreasoningThe annotators were not told the topic of the paper in order to minimize potential bias.
“The annotators are informed that their work will be included in a research paper, but the topic of the paper is not disclosed in order to minimize potential bias.”
Quote in the paperNot a substantial claimProbably not a substantial claim.
PremisecitationPrior work often assumes that LLMs are inherently unstable to prompt changes.
“To address this, prior work often assumes that LLMs are inherently unstable to prompt changes.”
Quote in the paperSubstantial claimProbably not a substantial claim. The paper may give no reason why it holds.
Premisenone givenAll existing methods attribute the sensitivity to model behavior.
“However, all existing methods attribute the sensitivity to model behavior.”
Quote in the paperNot a substantial claimProbably not a substantial claim.
Study design and conductreasoningDue to computational constraints, we evaluate each benchmark using only 12 prompt templates.
“Due to computational constraints, we evaluate each benchmark using only 12 prompt templates.”
Quote in the paperSubstantial claimProbably not a substantial claim. The paper’s conclusions may not depend on it.
PremisecitationMizrahi et al. (2024) conducted a large-scale study showing significant accuracy differences across prompt variants.
“Mizrahi et al. (2024) conducted a large-scale study showing significant accuracy differences across prompt variants.”
Quote in the paperNot a substantial claim0 of 1 sources checked- cites Mizrahi et al., 2024
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
Probably not a substantial claim.
Study design and conductnone givenAll evaluations use greedy decoding to ensure deterministic outputs.
“All evaluations use greedy decoding to ensure deterministic outputs.”
Quote in the paperSubstantial claim
The paper relies on this without support.
The seven evaluated LLMs are representative of modern LLMs.
“We evaluate LLaMA-3.1-8B-Instruct(LLaMA-3.1)(Dubey et al., 2024), Qwen2-7B-Instruct (Qwen-2) (Yang et al., 2024), Gemma-2-9B-it (Gemma-2) (Gemma Team et al., 2024), Ministral-8B-Instruct-v0.2 (Ministral) (MistralAI, 2024), GPT-4o-mini (July 2024), GPT-4.1-mini(April 2025), and Gemini 2.0 Flash (February 2025).”
The paper relies on this without support.
The six benchmarks are representative of the tasks used to evaluate LLMs.
“We evaluate on six benchmarks covering both multiple-choice and open-ended tasks.”
The paper relies on this without support.
Human annotations provide a reliable ground truth for answer correctness.
“To assess the reliability of LLM-judges, we compare them against human annotations.”
“Human annotators evaluate whether each answer matches the corresponding ground truth.”
The paper relies on this without support.
No official heuristic parser is currently available for SimpleQA, making rule-based evaluation infeasible.
“No official heuristic parser is currently available, making rule-based evaluation infeasible.”
Self-check
Percentages, sums and reported differences re-computed from the paper’s own numbers.
Each contradiction between two passages listed here was confirmed by an independent check.
Gemma-2 accuracy range is incompatible with its reported standard deviation
MajorAcross 12 prompts, an accuracy range of 0.17 requires a standard deviation of at least approximately 0.035, even if every other accuracy lies at the midpoint. The reported standard deviation of 0.005 cannot describe those same results.
“its accuracy varies by only 0.17 across the same set of prompt templates”
“with a much lower standard deviation of just 0.005.”
NarrativeQA heuristic rank correlation disagrees with Table 1
MajorThe results text gives the NarrativeQA heuristic correlation as 0.40, whereas Table 1 reports 0.5927. The difference materially changes the stated improvement to 0.87 under LLM-as-a-Judge.
“On NarrativeQA, the correlation rises from 0.40 (heuristics) to 0.87 (LLM-as-a-Judge).”
“| NarrativeQA † | 0.5927 | 0.8662 |”
Citations
Cites
A Coefficient of Agreement for Nominal Scales
Cohen, J. (1960) · Educational and Psychological Measurement · doi:10.1177/001316446002000104 · Cited by 42,398
Language Models are Few-Shot Learners
Brown, T. B., Mann, B., Ryder, N., Subbiah, M., Kaplan, J., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., et al. (2020) · arXiv · doi:10.48550/arxiv.2005.14165 · Cited by 23,372
Yang, A., Li, A., Yang, B., Zhang, B., Hui, B., Zheng, B., Yu, B., Gao, C., Huang, C., Lv, C., et al. (2025) · arXiv · doi:10.48550/arxiv.2505.09388 · Cited by 17,283
Measuring nominal scale agreement among many raters.
Fleiss, J. L. (1971) · Psychological Bulletin · doi:10.1037/h0031619 · Cited by 8,733
The proof and measurement of association between two things
Spearman, C. (1904) · The American Journal of Psychology · doi:10.2307/1412159 · Cited by 8,039
Judging LLM-as-a-Judge with MT-Bench and Chatbot Arena
Zheng, L., Chiang, W.-L., Sheng, Y., Zhuang, S., Wu, Z., Zhuang, Y., Lin, Z., Li, Z., Li, D., Xing, E. P., et al. (2023) · arXiv · doi:10.48550/arxiv.2306.05685 · Cited by 1,888
Can a Suit of Armor Conduct Electricity? A New Dataset for Open Book Question Answering
Mihaylov, T., Clark, P. E., Khot, T., & Sabharwal, A. (2018) · Conference on Empirical Methods in Natural Language Processing (EMNLP) · doi:10.18653/v1/d18-1260 · Cited by 1,602
Instruction Tuning for Large Language Models: A Survey
Zhang, S., Dong, L., Li, X., Zhang, S., Sun, X., Wang, S., Li, J., Hu, R., Zhang, T., Wang, G., et al. (2025) · ACM Computing Surveys · doi:10.1145/3777411 · Cited by 1,207
The Flan Collection: Designing Data and Methods for Effective Instruction TuningPreprint
Longpre, S., Hou, L., Vu, T., Webson, A., Chung, H. W., Tay, Y., Zhou, D., Le, Q. V., Zoph, B., Wei, J., et al. (2023) · International Conference on Machine Learning · doi:10.48550/arxiv.2301.13688 · Cited by 1,031
CoQA: A Conversational Question Answering Challenge
Reddy, S., Chen, D., & Manning, C. D. (2019) · Transactions of the Association for Computational Linguistics · doi:10.1162/tacl_a_00266 · Cited by 975
G-Eval: NLG Evaluation using Gpt-4 with Better Human Alignment
Liu, Y., Iter, D., Xu, Y., Wang, S., Xu, R., & Zhu, C. (2023) · Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing · doi:10.18653/v1/2023.emnlp-main.153 · Cited by 948
HellaSwag: Can a Machine Really Finish Your Sentence?
Zellers, R., Holtzman, A., Bisk, Y., Farhadi, A., & Choi, Y. (2019) · Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics · doi:10.18653/v1/p19-1472 · Cited by 830
Grattafiori, A., Dubey, A., Jauhri, A., Pandey, A., Kadian, A., Al-Dahle, A., Letman, A., Mathur, A., Schelten, A., Vaughan, A., et al. (2024) · arXiv · doi:10.48550/arxiv.2407.21783 · Cited by 505
Think you have Solved Question Answering? Try ARC, the AI2 Reasoning Challenge
Clark, P., Cowhey, I., Etzioni, O., Khot, T., Sabharwal, A., Schoenick, C., & Tafjord, O. (2018) · arXiv · doi:10.48550/arxiv.1803.05457 · Cited by 462
LongBench: A Bilingual, Multitask Benchmark for Long Context Understanding
Bai, Y., Lv, X., Zhang, J., Lyu, H., Tang, J., Huang, Z., Du, Z., Liu, X., Zeng, A., Hou, L., et al. (2024) · Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) · doi:10.18653/v1/2024.acl-long.172 · Cited by 400
Can large language models reason about medical questions?
Liévin, V., Hother, C. E., Motzfeldt, A. G., & Winther, O. (2024) · Patterns · doi:10.1016/j.patter.2024.100943 · Cited by 343
Measuring Mathematical Problem Solving With the MATH Dataset
Hendrycks, D., Burns, C., Kadavath, S., Arora, A., Basart, S., Tang, E., Song, D., & Steinhardt, J. (2021) · arXiv · doi:10.48550/arxiv.2103.03874 · Cited by 306
Measuring Massive Multitask Language Understanding
Hendrycks, D., Burns, C., Basart, S., Zou, A., Mazeika, M., Song, D., & Steinhardt, J. (2020) · arXiv · doi:10.48550/arxiv.2009.03300 · Cited by 291
Super-NaturalInstructions: Generalization via Declarative Instructions on 1600+ NLP Tasks
Wang, Y., Mishra, S., Alipoormolabashi, P., Kordi, Y., Mirzaei, A., Naik, A., Ashok, A., Dhanasekaran, A. S., Arunkumar, A., Stap, D., et al. (2022) · Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing · doi:10.18653/v1/2022.emnlp-main.340 · Cited by 247
Instruction Tuning for Large Language Models: A Survey
Zhang, S., Dong, L., Li, X., Zhang, S., Sun, X., Wang, S., Li, J., Hu, R., Zhang, T., Wu, F., et al. (2023) · arXiv · doi:10.48550/arxiv.2308.10792 · Cited by 241
Is chatGPT a good nlg evaluator? A preliminary study
Wang, J., Liang, Y., Meng, F., Sun, Z., Shi, H., Li, Z., Xu, J., Qu, J., & Zhou, J. (2023) · Proceedings of the 4th New Frontiers in Summarization Workshop · doi:10.18653/v1/2023.newsum-1.1 · Cited by 235
OpenAI, :, Hurst, A., Lerer, A., Goucher, A. P., Perelman, A., Ramesh, A., Clark, A., Ostrow, A., Welihinda, A., et al. (2024) · arXiv · Cited by 180
Gemma 2: Improving Open Language Models at a Practical Size
Gemma Team, Riviere, M., Pathak, S., Sessa, P. G., Hardin, C., Bhupatiraju, S., Hussenot, L., Mesnard, T., Shahriari, B., Ramé, A., et al. (2024) · arXiv · doi:10.48550/arxiv.2408.00118 · Cited by 149
Large Language Models Are State-of-the-Art Evaluators of Translation Quality
Kocmi, T., & Federmann, C. (2023) · arXiv · doi:10.48550/arxiv.2302.14520 · Cited by 122
State of What Art? A Call for Multi-Prompt LLM Evaluation
Mizrahi, M., Kaplan, G., Malkin, D., Dror, R., Shahaf, D., & Stanovsky, G. (2024) · Transactions of the Association for Computational Linguistics · doi:10.1162/tacl_a_00681 · Cited by 113
Mind Your Format: Towards Consistent Evaluation of In-Context Learning Improvements
Voronov, A., Wolf, L., & Ryabinin, M. (2024) · Annual Meeting of the Association for Computational Linguistics · doi:10.48550/arxiv.2401.06766 · Cited by 113
Efficient multi-prompt evaluation of LLMs
Bhardwaj, O., Choshen, L., de Oliveira, A. F. M., et al. (2024) · Neural Information Processing Systems · doi:10.48550/arxiv.2405.17202 · Cited by 100
Mind Your Format: Towards Consistent Evaluation of In-Context Learning Improvements
Voronov, A., Wolf, L., & Ryabinin, M. (2024) · Findings of the Association for Computational Linguistics ACL 2024 · doi:10.18653/v1/2024.findings-acl.375 · Cited by 77
Yang, A., Yang, B., Hui, B., Zheng, B., Yu, B., Zhou, C., Li, C., Li, C., Liu, D., Huang, F., et al. (2024) · arXiv · doi:10.48550/arxiv.2407.10671 · Cited by 63
Efficient multi-prompt evaluation of LLMs
Polo, F., Xu, R., Weber, L., Silva, M., Bhardwaj, O., Choshen, L., Oliveira, A. D., Sun, Y., & Yurochkin, M. (2024) · Advances in Neural Information Processing Systems 37 · doi:10.52202/079017-0707 · Cited by 60
Sclar, M., Choi, Y., Tsvetkov, Y., & Suhr, A. (2023) · arXiv · doi:10.48550/arxiv.2310.11324 · Cited by 52
POSIX: A Prompt Sensitivity Index For Large Language Models
Chatterjee, A., Renduchintala, H. S. V. N. S. K., Bhatia, S., & Chakraborty, T. (2024) · Findings of the Association for Computational Linguistics: EMNLP 2024 · doi:10.18653/v1/2024.findings-emnlp.852 · Cited by 34
GPQA: A Graduate-Level Google-Proof Q&A Benchmark
Rein, D., Hou, B. L., Stickland, A. C., Petty, J., Pang, R. Y., Dirani, J., Michael, J., & Bowman, S. R. (2023) · arXiv · doi:10.48550/arxiv.2311.12022 · Cited by 30
Hua, A., Dhaliwal, M. P., Pullela, L., Burke, R., & Qin, Y. (2024) · arXiv (Cornell University) · doi:10.48550/arxiv.2407.12843 · Cited by 20
Self-citationMeasuring short-form factuality in large language models
Wei, J., Karina, N., Chung, H. W., Jiao, Y. J., Papay, S., Glaese, A., Schulman, J., & Fedus, W. (2024) · arXiv · doi:10.48550/arxiv.2411.04368 · Cited by 14
POSIX: A Prompt Sensitivity Index For Large Language Models
Chatterjee, A., Renduchintala, H. S. V. N. S. K., Bhatia, S., & Chakraborty, T. (2024) · arXiv · doi:10.48550/arxiv.2410.02185
Flagged first, then by citations
Research works printed but not linked: 2
- Tomáš Kočiskỳ, Jonathan Schwarz, Phil Blunsom, Chris Dyer, Karl Moritz Hermann, Gábor Melis, and Ed- ward Grefenstette. 2018. The narrativeqa reading comprehension challenge. Transactions of the Asso- ciation for Computational Linguistics, 6:317-328.
- PMLR.
Cited by
Sorted by citations
Measuring Pragmatic Influence in LLM InstructionsPreprint
Geng, Y., Abend, O., Hovy, E. H., & Frermann, L. (2026) · arXiv (Cornell University) · doi:10.48550/arxiv.2602.21223 · Cited by 4
Where the Verifier Fails: A Category-Level Audit of Reward Signals in RLVRPreprint
Xin, E. (2026) · arXiv (Cornell University) · doi:10.48550/arxiv.2609.01354 · Cited by 3
Testing LLM Diagnostics in Endodontics: The Impact of Linguistic Variation on Unseen Cases
Batool, I., Naved, N., & Umer, F. (2026) · International Endodontic Journal · doi:10.1111/iej.70109 · Cited by 2
Lee, J., Koh, J., Tok, Z., Batra, H., & Clark, R. (2026) · Workshop on Generation, Evaluation and Metrics (GEM) · doi:10.18653/v1/2026.gem-main.23 · Cited by 1
Similar Pattern Annotation via Retrieval Knowledge for LLM-Based Test Code Fault LocalizationPreprint
Gharachorlu, G., Panahandeh, M., Briand, L., Gao, R., & Wan, R. (2026) · arXiv (Cornell University) · doi:10.48550/arxiv.2605.07957 · Cited by 1
Discourse-Role Labels as Presentation-Time Variables for Context Use in Language ModelsPreprint
Zhu, J. G., Li, X., & Liu, W. (2026) · arXiv (Cornell University) · doi:10.48550/arxiv.2606.04109 · Cited by 1
Measuring GEO Visibility: Prompt Corpora Define the Answer MarketPreprint
Martinez, O. (2026) · arXiv (Cornell University) · doi:10.48550/arxiv.2609.06811 · Cited by 1
Murugan, M., Yuan, B., Stephen, J. A., Gijavanekar, C., Xu, S., Kadirvel, S., Rivera-Munoz, E. A., Manita, V. M. C., Delca, F., Gibbs, R. A., et al. (2026) · medRxiv · doi:10.64898/2026.02.18.26346543
Prompt Engineering for Low-Resource Adaptation of Large Language Models in Agricultural Breeding
Yan, Z., Cao, Y., Gao, X., Chen, J., & Dong, Y. (2026) · International Conference on Big Data, Information and Computer Network (BDICN) · doi:10.1145/3801228.3801294
Burnat, F. A. D., & Davidson, B. I. (2026) · arXiv (Cornell University) · doi:10.48550/arxiv.2605.06327
Sajan, A. K. (2026) · Research Square · doi:10.21203/rs.3.rs-9960161/v1
Zhou, X. (2026) · Applied Psychological Measurement · doi:10.1177/01466216261471171
Azaïs, M.-A., Coustaty, M., & Guillaume, J. (2026) · Lecture notes in computer science · doi:10.1007/978-3-032-31438-3_34
Panossian, A. (2026) · arXiv (Cornell University) · doi:10.48550/arxiv.2608.10314
Sajan, A. K. (2026) · Research Square · doi:10.21203/rs.3.rs-10415268/v1
When Does Supervised Fine-Tuning Reduce Instruction Sensitivity?Preprint
Choi, J. (2026) · arXiv (Cornell University) · doi:10.48550/arxiv.2608.26661
Response Stability of Large Language Models Under Meaning-Preserving Prompt Variation
Idrees, & Liu, Y. (2026) · International Journal of Research and Innovation in Social Science · doi:10.47772/ijriss.2026.100800423
Towards Cognitive-Agent-Driven Prompt Optimization for Automated Program Repair
Gaia, F. N., & Maia, M. D. A. (2026) · ACM/IEEE International Conference on Software Engineering: Companion Proceedings (ICSE-Companion) · doi:10.1145/3774748.3787658
Martinez, O. (2026) · SSRN Electronic Journal · doi:10.2139/ssrn.7424318
Prompt Sensitivity or Evaluation Artifact? A Task-Aware Analysis for Large Language Models
Muthukumarana, S., Wijenayake, B., Godaliyadda, R., & Ekanayake, P. (2026) · Moratuwa Engineering Research Conference (MERCon) · doi:10.1109/mercon71835.2026.11691277
Boullineau, E., & Arciniegas, J. D. M. (2026) · SSRN Electronic Journal · doi:10.2139/ssrn.7490698
When Verifiable Counts Depend on Wording: Auditing Wording Robustness in Instruction FollowingPreprint
Zhan, Q., Jang, S., Dong, Z., Hu, M., Chen, Z., & Qu, T. (2026) · arXiv (Cornell University) · doi:10.48550/arxiv.2610.05278
Quantization Effects on Tool-Failure Recovery Vary Across Prompts and Evaluation DesignsPreprint
Hu, Y. (2026) · arXiv (Cornell University) · doi:10.48550/arxiv.2610.07781
When the Trigger Fails: Metamorphic Testing of Tool-Calling AI Agents
de Amorim, W. F. (2026) · Simpósio Brasileiro de Testes de Software Sistemático e Automatizado (SAST) · doi:10.5753/sast.2026.32354
SFT-as-Context Mitigates Forgetting in Supervised Fine-TuningPreprint
Tang, K., Hua, A., Qian, C., Tiwari, S., & Qin, Y. (2026) · arXiv (Cornell University) · doi:10.48550/arxiv.2610.11132
Self-citationLanguage Is a Security Surface for Agent GuardsPreprint
Demirkiran, H., & Ernstberger, J. (2026) · arXiv (Cornell University) · doi:10.48550/arxiv.2610.22244
Same Answer, New Box: How Much of RLVR's Out-of-Domain Gain Is Answer Format?Preprint
Lyu, Z. (2026) · arXiv (Cornell University) · doi:10.48550/arxiv.2610.20639
On the Fragility of Tasks under Composition: Implicit Budgeting in Language ModelsPreprint
Li, C., Brockmoeller, T., Jegelka, S., Gupta, S., & Karalias, N. (2026) · arXiv (Cornell University) · doi:10.48550/arxiv.2610.27190
Metadata
- Authors
- Hua, A., Tang, K., Gu, C., Gu, J., Wong, E., & Qin, Y.
- Affiliations
- University of California, Santa Barbara (US)
- University of California, Irvine (US)
- University of Pennsylvania (US)
- Venue
- Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing
- Published
- 2025
- Pages
- 19900-19910
- OpenAlex
- W4416036473
- Cited by
- 28
- References
- 14
- Citations by year
- 2026: 28
- FWCI
- 10.17
- Type
- conference paper
- Language
- en
- Open access
- hybrid
- doi.org
- Licence
- cc-by
- Locations
- Conference on Empirical Methods in Natural Language Processing (EMNLP) (publishedVersion, cc-by)
- arXiv (Cornell University) (submittedVersion)
- arXiv (Cornell University) (cc-by)
- Topics
- Physical Sciences › Computer Science › Artificial Intelligence › Topic Modeling
- Physical Sciences › Computer Science › Artificial Intelligence › Natural Language Processing Techniques
- Physical Sciences › Computer Science › Artificial Intelligence › Adversarial Robustness in Machine Learning
- Keywords
- prompt sensitivity, paraphrase, large language models, LLM evaluation, evaluation artifacts, LLM-as-a-judge, prompt templates
About these checks
Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.
A language model drew the claims from the paper on 2026-11-08, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).
A second model labelled each claim on 2026-11-08 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).
Cited sources are checked with a language model’s help (gpt-6-sol).
Statistics and arithmetic are recomputed, not judged.
Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.
What counts as a claim ›How we measure confidence ›How we check this ›
Are you an author, or do you see a problem with this page? Request a review
