acceptodds
Under review as a conference paper at ICLR 2027

Lost in Formulation: Disentangling and Measuring Prompt Sensitivity

Abstract

Language model outputs vary with both the phrasing of a prompt (prompt formulation) and the amount of information it provides about the intended question (prompt specificity). Separating these sources of sensitivity helps users decide how to rewrite a prompt, but existing prompt sensitivity metrics conflate them. We vary prompt formulation and prompt specificity systematically on questions from the AmbigQA dataset, collect repeated answers from three open-weight models, and compute ten existing metrics on these answers. A principal component analysis separates the metrics into three components, which we call correctness, formulation dependence, and output dispersion. While existing metrics capture correctness and output dispersion well, none captures formulation dependence without also counting decoding noise, the variation among repeated answers to the same prompt. We therefore introduce , an intraclass correlation of correctness across formulations that subtracts decoding noise. Although it does not enter the decomposition, tracks the formulation-dependence component more closely than existing metrics, also on held-out AmbigQA questions. Varying specificity and formulation confirms that the three components respond to different changes. Disambiguating a question raises correctness and lowers output dispersion but produces no detectable change in , whereas more varied formulations tend to raise but not the other two. Correctness, formulation dependence, and output dispersion are thus distinct parts of prompt sensitivity, and measures formulation dependence separately from decoding noise.

Primary area:
foundation or frontier models, including LLMs
Keywords:
prompt sensitivity, prompt robustness, paraphrase robustness, large language models, LLM evaluation, question ambiguity, intraclass correlation

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_MdawOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper explores prompt sensitivity of LLMs in generating answers. The authors put forward cases where slight changes in prompt formulation, specificity and sampling might drastically change the answer produced. They discuss three separate factors: correctness, formulation dependence and output dispersion to evaluate prompt sensitivity. They provide a new metric ρF\rho_F that measures how much the correctness of an LLMs answer depends on the prompt formulation. The authors experiment on the AmbigQA dataset showing the contribution of each of the factors on different aspects of prompt sensitivity.

Critical strengths
  1. The motivation feels decently well portrayed. With LLMs used in many parts of life, prompt sensitivity analysis is a relevant research problem.
  2. The authors perform a detailed analysis of ten existing metrics.
Critical weaknesses
  1. The experimentation and generalization of the findings is limited. The authors experiment only on the AmbigQA dataset which mainly contains short factual questions. Hence it can not be confidently established whether the claims hold true for reasoning or summarization tasks as well.
  2. Experiments have been performed only on the relatively small open-weight models.
  3. The proposed metric ρF\rho_F is conceptually similar to the existing metric ρu\rho_u. The contribution of this newly proposed metric seems to be somewhat incremental.
Questions for authors
  1. Is it possible to experiment on other datasets alongside AmbigQA, also on other tasks?
  2. Does the sensitivity values hold true for large weight models as well?
  3. Could the authors clarify what advantages ρF\rho_F provides over adapting existing variance-based metrics, particularly given that simpler metrics remain competitive in some of the experiments.
2 comments
AuthorsOpenReview ↗

Reply to Reviewer Mdaw

We thank you for the constructive review. Our general comment lists what we added, how AmbigQA answers are scored, and the results per model. We address your points in the following:

We added AMBROSIA, a text-to-SQL benchmark from which we use 600 questions, as a second dataset. As on AmbigQA, disambiguation raises accuracy on AMBROSIA (by .23 to .49) and lowers semantic entropy HsemH_\text{sem} in all seven models. More varied formulations raise the hierarchical estimate of ρF\rho_F on AMBROSIA in six of the seven models. They change accuracy by less than one percentage point, with intervals that include zero, and HsemH_\text{sem} by less than .03. Therefore, we replicated two results on a different task. Accuracy and output dispersion respond to disambiguation, and ρF\rho_F responds to more varied formulations. Reasoning and open-ended tasks such as summarization remain untested (Section 6). Our general comment reports how ρF\rho_F changes under disambiguation on both datasets. On AMBROSIA's ambiguous questions, 29 to 58 % of the variance between formulations is associated with which annotated interpretation the answers take, against 5 to 15 % after disambiguation. We report these results in Section 4.3 and in Appendix D.7.

We added four larger open-weight models and replicated the effect of disambiguation on accuracy and semantic entropy in them. Disambiguation raises accuracy and lowers HsemH_\text{sem} in all seven models from 7B to 72B on both datasets (table in our general comment). More varied formulations raise ρF\rho_F in three of the four larger models on each dataset. On AmbigQA, it falls in Gemma-4-31B (by .03), and on AMBROSIA, a rise is not detected in Qwen2.5-72B. The level of ρF\rho_F differs between models (mean .10 to .68 across the four larger models on AmbigQA). We report these runs in Section 4.3 and in Appendices D.2 and D.8.

We compared ρF\rho_F with ρu\rho_u on the answers of three models and found two differences. We agree that the two metrics are conceptually similar. Both are shares of variance between formulations, and both rank the models in the same order. First, the embedding variance share ρu\rho_u is computed on answer embeddings, not on correctness. In 25 to 36 % of AmbigQA cells, all answers are correct or all answers are incorrect, although the answers still differ. In these cells, ρu\rho_u averages .17 to .38. Second, ρu\rho_u does not subtract decoding noise. When we shuffle the answers of a cell across formulations, no formulation effect remains. ρu\rho_u then still averages .09 at ten formulations and ten samples, whereas the hierarchical estimate of ρF\rho_F stays within .01 of zero on AmbigQA. Therefore, ρF\rho_F complements ρu\rho_u. ρF\rho_F measures the share of the variation in correctness that depends on the formulation, net of decoding noise.

We tested the three properties that ρF\rho_F adds over simpler variance-based metrics, on the AmbigQA answers of the same three models. ρF\rho_F is computed on correctness, it is close to zero when the formulation has no effect, and it depends less on the number of formulations and samples. At a fixed number of formulations and samples, ρF\rho_F ranks cells in the same way as the plain between-formulation share η2\eta^2 (ρu\rho_u adapted to correctness). When we shuffle outcomes across formulations, η2\eta^2 averages .47 at two samples per formulation, whereas the method-of-moments estimate of ρF\rho_F, before truncation at zero, stays within .01 of zero. From two to twenty samples per formulation, the hierarchical estimate of ρF\rho_F falls by 4 to 20 %, whereas ρu\rho_u falls by 31 to 68 % on the same cells. From four to fifty formulations, ρF\rho_F stays within .04, whereas the spread rises by .16 to .20. Therefore, we recommend ρF\rho_F to measure how much the correctness of an answer depends on the formulation. To locate the questions where a change of formulation gains most, we recommend the variance between formulations σB2\sigma_B^2 or the sampled spread. Section 4.3 states this comparison, and Table 10 in Appendix D.4 compares the three metrics criterion by criterion.

Please let us know if anything remains unclear.

Thank you again for your thoughtful review. If you have a chance to look at our revision and response, we’d welcome your thoughts on whether they address your concerns.

Reviewer_rwjTOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper investigates the sensitivity of large language models (LLMs) to prompts, categorizing this phenomenon into three distinct dimensions. Through an analysis of existing evaluation metrics, the authors find that these metrics struggle to effectively disentangle the impact of prompt phrasing from the inherent stochastic noise in the model's generation process. To address this, they propose a new metric, ρF\rho_F, based on the Intraclass Correlation Coefficient (ICC), designed to quantify the influence of prompt phrasing on the correctness of model responses. Experiments conducted on the AmbigQA dataset using three open-source language models reveal that increasing the specificity of prompts enhances response accuracy and reduces output variance, while greater diversity in prompt phrasing tends to increase the model's reliance on specific wording. Overall, this research offers a more systematic approach to analyzing and evaluating prompt sensitivity; however, as the experiments are primarily limited to smaller models and relatively simple tasks, the generalizability of these findings warrants further validation.

Critical strengths
  • Based on ICC, a new method for measuring prompt sensitivity is proposed.
  • A clearly defined problem and robust statistical analysis.
  • The design can allow for a more targeted observation of how different factors alter metrics.
Critical weaknesses
  • ICC is derived from established statistical theory, and the hierarchical Beta-Binomial model is likewise not a new method. Therefore, the core innovation of this paper does not lie in a novel algorithm or theory; rather, it consists of applying the well-known concept of variance decomposition to the study of "prompt sensitivity" and providing an empirical validation of this approach.

  • The author only tested three 7B–8B models. The evaluation did not include more recent, larger-scale LLMs, reasoning-focused models, or mainstream closed-source models. Furthermore, tasks such as code generation, mathematical reasoning, and complex instruction following were not covered.

  • The paper employs a DeBERTa-v3-large NLI model to carry out several critical steps. There is a lack of human verification or validation.

  • The author emphasizes that ρF can measure formulation dependence at the per-question level. However, the results for split-half reliability appear quite unstable. It would be advisable to increase the number of formulations—for instance, to 20, 30, or 50—and plot how reliability changes with the sample size, thereby demonstrating the number of prompts required to achieve an acceptable level of stability.

Questions for authors

Refer to weakness.

3 comments
AuthorsOpenReview ↗

Reply to Reviewer rwjT

We thank you for the detailed review. Our general comment lists what we added, how AmbigQA answers are scored, and the results per model. We address your points in the following:

We clarified our contribution: we transfer the intraclass correlation (ICC) to prompt sensitivity and obtain ρF\rho_F, a metric of formulation dependence. ρF\rho_F is the standard one-way ICC of the correctness of the answers across the formulations of one question (Section 4.1). None of the ten existing metrics we evaluate separates the effect of the formulation on correctness from decoding noise. To our knowledge, the formulation's share of the variation in correctness has not been reported per question across meaning-preserving formulations. We also show with split-half reliability how many formulations ρF\rho_F needs (table below). We measure how ρF\rho_F responds when we vary the specificity of the question and the variety of the formulations (two datasets, seven models each). Section 2 compares ρF\rho_F with the closest existing work.

We added four larger open-weight models and replicated the effect of disambiguation on accuracy and semantic entropy in them. This effect is present in all seven models, from 7B to 72B, on both datasets (table in our general comment). To cover reasoning-focused models, we also ran Qwen3.8-27B with thinking enabled on the 500 AmbigQA questions. In this run, disambiguation again raises accuracy (+.28) and lowers semantic entropy HsemH_\text{sem} (-.18), and ρF\rho_F falls by .025 in the hierarchical estimate (95 % interval -.038 to -.013). More varied formulations raise ρF\rho_F in three of the four larger models on each dataset. On AmbigQA, it falls in Gemma-4-31B (by .03). In Qwen2.5-72B, Llama-3.1-70B, and Qwen3.8-27B, it rises by .01 to .02, which is of the order of the variation between repeated runs (about .01). On AMBROSIA, a rise is not detected in Qwen2.5-72B. We did not test closed models because reproducing our answers requires fixed weights and a fixed sampler. We report these runs in Section 4.3 and in Appendices D.2, D.8, and D.9.

We added AMBROSIA, a text-to-SQL benchmark, as a second dataset and replicated the finding on correctness and output dispersion on it. As on AmbigQA, disambiguation raises accuracy (by .23 to .49) and lowers HsemH_\text{sem} on AMBROSIA in all seven models. On AMBROSIA, more varied formulations raise the hierarchical estimate of ρF\rho_F in six of the seven models (Section 4.3 and Appendix D.7). We acknowledge in the limitations (Section 6) that reasoning, programming, and open-ended tasks remain untested.

We conducted a human validation of the three NLI steps. For the answer scoring, we compared the NLI rule in which the intended answer must entail the answer with a language model as judge, among 15 candidate rules. We score with a model as a judge (Phi-4), because it agrees better with the human labels (κ = .88 against .70). It also finds more of the correct answers (recall .94 against .73, precision .94 against .95). NLI checks that a formulation asks the same question and clusters the answers for semantic entropy. Both steps test entailment in both directions, as existing work on semantic uncertainty does (Kuhn et al., 2023; Farquhar et al., 2024). Of the first ten formulations of each generator, 96 to 99 % ask the same question as the original. The NLI model joins 34 % of the pairs with the same answer. Therefore, HsemH_\text{sem} overstates output dispersion on AmbigQA. On AMBROSIA, we define clusters by equal execution results, without an NLI model, and disambiguation also lowers HsemH_\text{sem} there. Appendix D.6 reports the agreement of every candidate rule (Table 15) and the main results under both rules as a sensitivity analysis (Table 16).

We generated up to 50 formulations per question and measured how the reliability of single-question values of ρF\rho_F changes with the number of formulations. The table shows the reliability on all 500 AmbigQA questions for all seven models, as the correlation between two disjoint sets of n formulations. Two disjoint sets drawn from 50 formulations hold at most 25 each, so the last column steps the value at 25 up to 50 with the Spearman-Brown formula.

Modeln = 10n = 20n = 25Projected to 50
Qwen2.5-7B.53.63.67.80
Llama-3.1-8B.53.65.68.81
Mistral-7B.60.73.77.87
Qwen2.5-72B.61.72.76.86
Llama-3.1-70B.54.64.67.80
Gemma-4-31B.64.68.68no projection
Qwen3.8-27B.42.56.60.75

For Gemma-4-31B, the measured reliability does not rise from 20 to 25 formulations, so we do not project it to 50. For Qwen2.5-7B, Llama-3.1-8B, and Mistral-7B, a single question needs about 15 to 29 formulations for a reliability of .70 and 30 to 50 for .80. Figure 6 and Table 13 in Appendix D.5 show how reliability changes with the number of formulations.

We are glad to clarify any remaining points.

Thanks for your response and new experiments results. Most of my concerns have been resolved. I would like to improve my rate to weak accept.

Thank you for taking the time to consider our response and the new results. We appreciate your feedback!

Reviewer_dspZOpenReview ↗
Rating
3(+1)
Confidence
1
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper investigates prompt sensitivity by varying prompt specificity and prompt formulation on AmbigQA across three open-weight models. The authors interpret a three-component solution from principal component analysis of ten existing metrics as correctness, formulation dependence, and output dispersion. They introduce ρF\rho_F, a metric based on intraclass correlation (ICC) of correctness that subtracts decoding noise, and evaluate it on held-out AmbigQA questions and by varying specificity and formulation.

Critical strengths
  1. The paper introduces ρF\rho_F to measure formulation dependence separately from decoding noise, complementing accuracy and semantic entropy.
  2. The evaluation combines component analysis, tests on held-out questions, permutation checks, tests that vary specificity and formulation, and split-half reliability analysis.
  3. The paper openly acknowledges its limitations, including limited reliability of single-cell values, reliance on an NLI model for scoring and formulation checks, and restricted experimental scope.
Critical weaknesses
  1. The methodological contribution beyond standard ICC and prior correctness-based variance decomposition needs clearer justification. Support for the three-factor interpretation is sensitive to metric selection: parallel analysis retains two components for the ten existing metrics and three after adding ρF\rho_F.
  2. The method-of-moments (MoM) estimate has limited coverage (45–66% of cells) and modest split-half reliability (0.38–0.57) in the main experiment, limiting confidence in single-cell estimates.
Questions for authors
  1. Does parallel analysis retain three components with hierarchical rather than MoM estimates of ρF\rho_F?

  2. Given the results in Table 6 with up to ten formulations per cell (line 435), can you report MoM coverage and split-half reliability for different formulation counts?

3 comments
AuthorsOpenReview ↗

Reply to Reviewer dspZ

We thank you for the thorough review. Our general comment lists what we added and how AmbigQA answers are scored. We address your points in the following reply:

We clarified the contribution beyond the standard intraclass correlation (ICC). Our contribution is to measure formulation dependence. We transfer the standard one-way ICC to the correctness of the answers to one question and obtain ρF\rho_F, the share of the variation in correctness that is due to the formulation, net of decoding noise. None of the ten existing metrics separates the effect of the formulation on correctness from decoding noise. We show how many formulations ρF\rho_F needs: ten are sufficient to compare models and question sets, whereas the value of a single question needs about 15 to 29 for a reliability of .70. We also measure how ρF\rho_F responds when we vary the specificity of the question and the variety of the formulations. Section 2 and Appendix C compare ρF\rho_F with the closest existing work.

We quantified with a subset analysis how the three-factor interpretation depends on the metric selection. We agree that the number of components depends on which metrics are included. We specified this analysis after your review. We counted the retained components for all 312 subsets of the eleven metrics that keep at least two metrics per factor and include ρF\rho_F. On AmbigQA, 64 % of the subsets retain three components on the main questions. The revised paper does not rest the three factors on the number of components alone. Section 4.3 tests how their metrics respond when we vary specificity and formulation. For ρF\rho_F on AmbigQA, the tests detect no change under disambiguation in Qwen2.5-7B and Llama-3.1-8B. In Mistral-7B, they detect a rise that is significant after one of the two corrections for multiple testing (p = .031 after Benjamini-Hochberg and p = .073 after Holm). More varied formulations raise ρF\rho_F in six of the seven models on each dataset and change accuracy by at most .02. In the original three models of the submission, the hierarchical estimate rises by .02 to .05. We report these tests in Appendices D.1 and D.2.

We repeated the parallel analysis with the hierarchical estimate for the three models of the submission, and it retains three components (your question 1). The parallel analysis retains three components on all 500 AmbigQA questions (third eigenvalue 1.41 against a threshold of 1.16), in every posterior draw. It also retains three components on the initially submitted 150 questions (1.50 against 1.29) and on the 100 held-out questions (1.50 against 1.36). As a control, we replaced ρF\rho_F with a copy of ρu\rho_u. Parallel analysis then also retains three components (1.48 against 1.16). The number of retained components shows that formulation dependence forms a third component on AmbigQA once a formulation dependence metric is added. Table 4 in Appendix B shows the result for every estimate.

We computed coverage and split-half reliability for four to 50 formulations (your question 2). We computed all values below on the 500 questions, scored with the judge, for the three models of the submission. Coverage is the share of cells in which the method-of-moments estimate is defined. A cell is one question at one specificity level for one model. In subsets of the ten formulations of the main experiment, coverage is .34 to .62 at four formulations and .44 to .74 at ten. Split-half reliability is .25 to .32 at four formulations and .56 to .64 at ten. In new sets of up to 50 formulations, coverage is .64 to .89 at 50. We measured the correlation between two disjoint sets of 25 formulations (.67 to .77). The Spearman-Brown formula steps it up to .80 to .87 for all 50 formulations (the highest measured value is .77). Coverage is limited because all answers are correct or all answers are incorrect in 27 to 56 % of AmbigQA cells at ten formulations. Therefore, we use the hierarchical estimate, which is defined for every cell, in all analyses that require a value for every cell.

We agree that ten formulations do not support values for single cells. The revised paper states this limit. Section 4.3 reads: "Split-half reliability on AmbigQA is 0.60/0.56/0.64 at ten formulations, which leaves single-cell values too noisy to report on their own, whereas the order of the models holds in 100 % of question bootstraps." For these three models, a single question needs about 15 to 29 formulations for a reliability of .70. This number depends on the model and on whether the value is measured or projected. Appendix D.5 has the full tables and a figure (Tables 12 and 13, Figure 6). Our reply to Reviewer rwjT lists the measured reliability per model.

We are happy to answer any further questions.

The additional experiments and revisions in the rebuttal have addressed my questions. Although the method novelty beyond the standard one-way ICC remains modest, the positioning is now clear and honest, so I am willing to raise my score.

Thank you for taking the time to read the revision. We appreciate your feedback!

Comments on the paper

1 comment
AuthorsOpenReview ↗

General comment: summary of the revision and results per model

We thank the three reviewers for their thoughtful and constructive reviews, and we appreciate that they highlighted the value of systematically analyzing prompt sensitivity, including our comparison of existing metrics and our approach to separating the effect of prompt formulation on answer correctness from decoding noise. To address the reviewers’ feedback, we implemented the following in our work:

  • We extended AmbigQA from 150 to 500 questions, with 100 additional held-out questions for the component analysis (Reviewers rwjT and Mdaw).
  • We added a second dataset, AMBROSIA (Reviewers rwjT and Mdaw). We used 600 questions of this text-to-SQL benchmark (Saparina & Lapata, 2024), 500 for the analyses and 100 for validation. An answer counts as correct when its result matches that of the intended query.
  • We added four larger models (Reviewers rwjT and Mdaw). We added Qwen2.5-72B, Llama-3.1-70B, Gemma-4-31B, and Qwen3.8-27B to the three models of the original submission (Qwen2.5-7B, Llama-3.1-8B, and Mistral-7B).
  • We generated up to 50 formulations per question for all 500 AmbigQA questions (Reviewers dspZ and rwjT). Ten formulations are sufficient to compare models and question sets, whereas single questions need more formulations to ensure robustness.
  • We conducted a human validation of the three natural language inference (NLI) steps (Reviewer rwjT). On 512 AmbigQA answers with human labels, a language model as judge (Phi-4) agrees better with the labels than the NLI rule (κ = .88 against .70). The revised paper therefore scores AmbigQA answers with the judge. NLI still checks the formulations and clusters the answers for semantic entropy. A second author independently labeled a subset (κ = .85).
  • We compared ρF\rho_F with ρu\rho_u, η2\eta^2, and the spread of accuracy across formulations (Reviewers dspZ and Mdaw). Of these metrics, ρF\rho_F is the only one that both is computed on the correctness of the answers and is close to zero when the formulation has no effect.
  • We repeated the component analysis with additional controls (Reviewer dspZ). The ten existing metrics yield three components: correctness, output dispersion, and formulation dependence. In those metrics, the representation of the formulation dependence component is weak, because its two existing metrics have low agreement (r = .46) and neither measures correctness net of decoding noise (eigenvalue 1.07 on AmbigQA and 1.13 on AMBROSIA against a threshold of 1.14). ρF\rho_F gives this component a metric. With ρF\rho_F, parallel analysis retains three components on five of the six sets of questions and models (Appendix B).
  • We clarified how ρF\rho_F relates to existing work (Reviewers dspZ and rwjT). ρF\rho_F is the standard one-way intraclass correlation (ICC), applied to the correctness of the answers to one question. The closest papers are Cox et al. (2025), Feng et al. (2026), and Mustahsan et al. (2025).

The table summarizes the main results for all seven models on both datasets (AmbigQA / AMBROSIA). It shows the change of accuracy and HsemH_\text{sem} from the ambiguous to the disambiguated question and the change of ρF\rho_F (hierarchical estimate) from the narrow to the wide formulation generator.

ModelAccuracy, disambiguationHsemH_\text{sem}, disambiguationρF\rho_F, varied formulations
Qwen2.5-7B+.21 / +.35-.14 / -.11+.054 / +.025
Llama-3.1-8B+.24 / +.23-.37 / -.16+.023 / +.019
Mistral-7B+.26 / +.24-.17 / -.13+.046 / +.036
Qwen2.5-72B+.27 / +.46-.08 / -.22+.017 / +.010
Llama-3.1-70B+.28 / +.41-.32 / -.32+.011 / +.026
Gemma-4-31B+.22 / +.49-.17 / -.06-.033 / +.054
Qwen3.8-27B+.24 / +.44-.18 / -.31+.020 / +.063
Mean of the seven models+.24 / +.37-.20 / -.19+.020 / +.033

Disambiguating a question raises accuracy in all seven models on both datasets, on average by 24 points on AmbigQA and 37 on AMBROSIA. Semantic entropy HsemH_\text{sem} falls in all seven models, on average by .20 and .19 bits on the two datasets. More varied formulations raise ρF\rho_F in six of the seven models on each dataset. On AmbigQA, it falls in Gemma-4-31B, and on AMBROSIA, a rise is not detected in Qwen2.5-72B. Under more varied formulations, accuracy and HsemH_\text{sem} stay nearly flat (changes of at most .02 and .04 bits). Under disambiguation, ρF\rho_F stays within .03 in every model on AmbigQA (-.001 on average). On AMBROSIA, it falls in six of the seven models, by .03 on average (Table 2). The revised PDF accompanies these replies.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.