CoChat
Sign Up
Cited by 34Topic ModelingNatural Language Processing TechniquesAdversarial Robustness in Machine Learning

POSIX: A Prompt Sensitivity Index For Large Language Models

Open access
Chatterjee, Anwoy, Renduchintala, H S V N S Kowndinya, Bhatia, Sumit, Chakraborty, Tanmoy·2024·Findings of the Association for Computational Linguistics: EMNLP 2024
DOI 10.18653/v1/2024.findings-emnlp.852
Open this paper in CoChat — free accountOpen PDF

Abstract

Despite their remarkable capabilities, Large Language Models (LLMs) are found to be surprisingly sensitive to minor variations in prompts, often generating significantly divergent outputs in response to minor variations in the prompts, such as spelling errors, alteration of wording or the prompt template.However, while assessing the quality of an LLM, the focus often tends to be solely on its performance on downstream tasks, while very little to no attention is paid to prompt sensitivity.To fill this gap, we propose POSIX -a novel PrOmpt Sensitivity IndeX as a reliable measure of prompt sensitivity, thereby offering a more comprehensive evaluation of LLM performance.The key idea behind POSIX is to capture the relative change in loglikelihood of a given response upon replacing the corresponding prompt with a different intent-preserving prompt.We provide thorough empirical evidence demonstrating the efficacy of POSIX in capturing prompt sensitivity and subsequently use it to measure and thereby compare prompt sensitivity of various opensource LLMs.We find that merely increasing the parameter count or instruction tuning does not necessarily reduce prompt sensitivity whereas adding some few-shot exemplars, even just one, almost always leads to significant decrease in prompt sensitivity.We also find that alterations to prompt template lead to the highest sensitivity in the case of MCQtype tasks, whereas paraphrasing results in the highest sensitivity in open-ended generation tasks.The code for reproducing our results is open-sourced at https://github.com/kowndinya- renduchintala/POSIX.

AI snapshot

Written from the paper’s full text.

TL;DR

The paper introduces POSIX, a prompt sensitivity index that measures relative changes in an LLM’s log-likelihood for a response when a prompt is replaced by an intent-preserving variant. Using POSIX across MMLU, Alpaca, and BBH, the authors find that instruction tuning and larger parameter counts do not necessarily reduce prompt sensitivity, while even one few-shot exemplar usually does.

Key findings

  • POSIX empirically correlates with four intended factors: response diversity, response distribution entropy, semantic coherence, and variance in confidence; the paper reports positive correlation for unique responses, response distribution entropy, and log-probability variance, and negative correlation for average cosine similarity.
  • Instruction tuning does not necessarily reduce sensitivity: chat/instruct models are generally less sensitive than base models for template variations on MMLU, with Mistral as the only exception, but instruct versions tend to be more sensitive in other variation categories, especially on Alpaca. FLAN-only fine-tuned models are mostly worse than chat versions except for prompt templates on MMLU, where FLAN-only models significantly outperform chat versions.
  • Adding few-shot exemplars, even just one, almost always leads to a significant decrease in prompt sensitivity; additional exemplars yield diminishing gains, with sensitivity staying about the same or slightly decreasing.
  • Prompt template is the most sensitive variation type for MCQ-type tasks, while paraphrasing is almost always the most sensitive variation type for open-ended generation on Alpaca, with OLMo being the only exception. Model scale results show OLMo 7B is significantly more sensitive than OLMo 1B on MMLU but comparable on Alpaca, and Llama-2 13B is not guaranteed to be less sensitive than Llama-2 7B.

Methods

  • POSIX is defined from an intent-aligned prompt set \(X=\{x_i\}_{i=1}^N\) and responses \(Y=\{y_i\}_{i=1}^N\): \(\psi_{M,X}=\frac{1}{N(N-1)}\sum_{i=1}^N\sum_{j=1}^N \frac{1}{L_{y_j}}\left|\log\frac{P_M(y_j|x_i)}{P_M(y_j|x_j)}\right|\), where \(L_{y_j}\) is response length. POSIX averages \(\psi_{M,X}\) over a dataset of intent-aligned prompt sets.
  • Experiments use MMLU for classification as MCQ, Alpaca for open-ended generation, and BBH families; models include LLaMA-2 7B base/chat, LLaMA-3 8B base/instruct, Mistral 7B base/instruct, OLMo 7B base/instruct, plus OLMo 1B, Llama-2 13B, and FLAN-fine-tuned LLaMA-2 7B and Mistral 7B. All experiments ran on 8 NVIDIA A100-SXM4-80GB GPUs.
  • For each prompt, 60 intent-aligned variants are generated: 20 spelling-error variants, 20 prompt-template variants, and 20 GPT-3.5-Turbo paraphrases. POSIX is computed per variation type using the original prompt plus its 20 corresponding variants; the mixture category uses 21 variants sampled uniformly from the three categories. MMLU generates 5 tokens, Alpaca 30 tokens. BBH uses only prompt-template variation because many BBH tasks do not preserve intent under spelling errors or paraphrasing.

Limitations

  • The paper states that POSIX has high computational complexity: it needs \(O(MN^2)\) log-likelihood comparisons if \(M\) is the total number of prompts and \(N\) is the number of variations per prompt.
  • The paper notes an ethical/data limitation: while paraphrasing MMLU prompts with GPT-3.5-Turbo, quite a few prompts were flagged as violent or biased, mostly from the moral_scenarios split, and these were removed from analyses.
  • Evident from the text: the empirical evaluation is limited to open-source models and specific datasets (MMLU, Alpaca, BBH), and BBH is restricted to template variations only because other perturbations were not intent-preserving for many tasks.

Model-written summary. Check it against the paper before relying on it.

Integrity signals

Registry

Crossref · 10.18653/v1/2024.findings-emnlp.852

Cites

0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Claims the paper makes

2 of 5 cited sources checked · 0 supported · 0 flagged · accuracy not yet measured

  • Premisecitation

    Large Language Models are sensitive to minor intent-preserving prompt variations, often producing significantly divergent outputs in response to such variations.

    “Despite their remarkable capabilities, Large Language Models (LLMs) are found to be surprisingly sensitive to minor variations in prompts, often generating significantly divergent outputs in response to minor variations in the prompts, such as spelling errors, alteration of wording or the prompt template.”
    • Also stated in: § 1 Introduction · plain
      “However, LLMs have been found to be surprisingly sensitive even to the smallest of variations in prompts that do not significantly alter its meaning -such as wording, prompt template or even minor spelling errors -so much so that prompt engineering, which is a process of iteratively tuning prompts to elicit desired responses, has become a widespread practice (Reynolds and McDonell, 2021).”
    Quote in the paperSubstantial claim1 of 1 sources checked
    • cites Reynolds and McDonell, 2021 source

      Section 4 (p. 4) says rewriting a prompt can significantly change task performance, and §4.5 (p. 6) describes how small formatting and wording changes affect whether GPT-3 translates a sentence or continues in French. The paper supports sensitivity to some prompt variations, but the cited passages do not establish that divergent outputs occur often or address spelling errors specifically. (LLM, conf=0.85)

      Experimental — accuracy not yet measured
  • Premisecitation

    Standard evaluation benchmarks focus predominantly on performance metrics like exact match, leaving prompt sensitivity sidelined.

    “Despite prompt sensitivity being a crucial aspect for assessing the usability of an LLM, standard evaluation benchmarks such as MMLU (Hendrycks et al., 2021) or BBH (Suzgun et al., 2022) focus predominantly on performance metrics like exact match, leaving prompt sensitivity sidelined.”
    Quote in the paperSubstantial claim1 of 2 sources checked
    • cites Hendrycks et al., 2021

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

    • cites Suzgun et al., 2022 source

      The Evaluation protocol states that BBH measures accuracy by exact match, supporting the performance-metric part of the claim. But §3 (Experimental Setup) compares answer-only and chain-of-thought prompting, and §4.1 reports that prompt instructions and answer options affect results. The paper does not establish that prompt sensitivity is wholly sidelined. (LLM, conf=0.83)

      Experimental — accuracy not yet measured
  • Findingown data

    Adding few-shot exemplars, even just one, almost always leads to a significant decrease in prompt sensitivity.

    “We find that merely increasing the parameter count or instruction tuning does not necessarily reduce prompt sensitivity whereas adding some few-shot exemplars, even just one, almost always leads to significant decrease in prompt sensitivity.”
    • Also stated in: § 5.4 Few-shot Exemplars and Sensitivity · plain
      “The key finding is that adding few-shot exemplars, even if it just a single example can significantly boost the robustness of LLMs towards variations in prompts.”
    Quote in the paperSubstantial claim
  • Findingown data

    Merely increasing parameter count or instruction tuning does not necessarily reduce prompt sensitivity.

    “We find that merely increasing the parameter count or instruction tuning does not necessarily reduce prompt sensitivity whereas adding some few-shot exemplars, even just one, almost always leads to significant decrease in prompt sensitivity.”
    • Also stated in: § 5.2 Effect of Instruction Tuning on Sensitivity · hedged
      “Note that this implies that instruction tuning does not necessarily improve model sensitivity.”
    Quote in the paperSubstantial claim
  • Findingown data

    Prompt template is the most sensitive variation type for multiple-choice question (MCQ) tasks.

    “From Table 1, it can be observed that prompt template is the most sensitive variation type in the case of MCQs”
    • Also stated in: § 6 Conclusions · plain
      “prompt template is the most sensitive variant type for MCQ tasks”
    Quote in the paperSubstantial claim
  • May combine two claims. 14
  • Classified as probably not a substantial claim. 7

  • May combine two claims.

    Findingown data

    Alterations to prompt template lead to the highest sensitivity in MCQ-type tasks, whereas paraphrasing results in the highest sensitivity in open-ended generation tasks.

    “We reveal an interesting observation based on POSIX computation -variations in template bring about maximum sensitivity in the case of MCQ-type tasks while variations in wording lead to maximum effect in the case of open-ended generation tasks (Section 5).”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    POSIX incorporates all four factors: response diversity, response distribution entropy, semantic coherence, and variance in confidence.

    “We empirically show in Section 5.1 that POSIX incorporates all the four factors listed above.”
    • Also stated in: § 5.1 Evaluating the Efficacy of POSIX · boosted
      “Thus, the plots serve as an empirical validation for the fact that POSIX incorporates each of the factors described in Section 3.”
    Quote in the paperSubstantial claim
    • The authors state this as an assumption: For an LLM that is not sensitive to prompt variations, the probability of generating a given response remains approximately the same when the prompt is replaced by an intent-preserving variant, i.e., P(y1|x1) ≈ P(y1|x2).

  • May combine two claims.

    Findingreasoning

    Adding more few-shot examples might yield diminishing gains; the additional value of a second or third few-shot exemplar is not much, and prompt sensitivity either remains about the same or slightly decreases.

    “Although, adding even more few-shot examples might yield diminishing gains, i.e., when compared to the value that a single example adds, the additional value of a second or third few-shot exemplar is not that much -prompt sensitivity either remains about the same or slightly decreases.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    Mistral models have the least divergence in sensitivity between base and instruct variants, whereas OLMo models have the most disparity.

    “Furthermore, the Mistral models seem to have the least divergence in sensitivity between the base and instruct variants whereas the OLMo models have the most disparity.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Findingown data

    Most of the prompts flagged as violent or biased during paraphrasing of MMLU were from the moral_scenarios split of MMLU.

    “Most of them were from the moral_scenarios split of MMLU.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    The number of unique responses, response distribution entropy, and log-probability variance when all responses are identical have positive correlation with POSIX, while average cosine similarity between responses is negatively correlated with POSIX.

    “The number of unique responses, the response distribution entropy and the log-probability variance (in case all responses are identical) -all have positive correlation with POSIX and the average cosine similarity between the responses is negatively correlated with POSIX.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    The 7B OLMo model is significantly more sensitive than the 1B model on MMLU, but they are comparable on Alpaca.

    “We observe that the 7B model is significantly more sensitive compared to the 1B model on the MMLU dataset; however, they are comparable in the case of the Alpaca dataset, in the case of OLMo.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingreasoning

    In most cases, the chat version is better than FLAN-only models in terms of sensitivity, except for prompt template on MMLU, where FLAN-only models significantly outperform the chat versions.

    “In most cases, the chat version is better than FLAN-only models in terms of sensitivity, except in case of prompt template for MMLU, where FLAN-only models significantly outperform the chat versions.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S4 (based).

    Interpretationnone given

    Models might show significant variance in loglikelihoods of responses for different prompt variations even when the responses are identical.

    “Furthermore, models might show significant variance in the loglikelihoods of responses for different prompt variations even when the responses are identical.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Premisecitation

    Existing studies are limited to deterministic responses like multiple-choice questions or short answers and do not consider sensitivity in open-ended generative tasks.

    “Finally, the existing studies are also limited to deterministic responses, like multiple-choice questions or short answers (Kwiatkowski et al., 2019) and do not consider the sensitivity in the case of open-ended generative tasks.”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites Kwiatkowski et al., 2019 source

  • May combine two claims.

    Interpretationown data

    POSIX is a novel prompt sensitivity index that reliably measures prompt sensitivity and offers a more comprehensive evaluation of LLM performance.

    “To fill this gap, we propose POSIX -a novel PrOmpt Sensitivity IndeX as a reliable measure of prompt sensitivity, thereby offering a more comprehensive evaluation of LLM performance.”
    • Also stated in: § 1 Introduction · plain
      “We introduce POSIX -a novel prompt sensitivity index that can act as a reliable measure of sensitivity of LLMs towards intent-preserving variations of prompts (Section 3).”
    Quote in the paperSubstantial claim
    • The authors state this as an assumption: For an LLM that is not sensitive to prompt variations, the probability of generating a given response remains approximately the same when the prompt is replaced by an intent-preserving variant, i.e., P(y1|x1) ≈ P(y1|x2).
    • Rests on an assumption: A real-world user may not always be able to formulate the optimal prompt every time.
    • Rests on an assumption: A universally applicable optimal prompt that can work across different model architectures may not exist.
    • Rests on an assumption: Prompt sensitivity is a crucial aspect for assessing the usability of an LLM.

  • May combine two claims.

    Interpretationown data

    Accuracy and sensitivity are two separate aspects; higher accuracy does not necessarily imply better sensitivity and vice-versa.

    “This only re-emphasizes the fact that accuracy and sensitivity are two separate aspects -and higher accuracy does not necessarily imply better sensitivity and vice-versa.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Interpretationown data

    POSIX is a reliable and effective measure of the sensitivity of large language models to intent-preserving variations in prompts.

    “We introduced POSIX -a novel prompt sensitivity index, as a reliable measure of sensitivity of LLMs towards intent-preserving variations in prompts such as spelling errors, prompt templates, and alterations in the wording.”
    • Also stated in: § 6 Conclusions · plain
      “We presented thorough empirical analysis for the efficacy of POSIX in capturing prompt sensitivity”
    • Also stated in: § G Efficacy of POSIX for Open-ended Generation · boosted
      “Figure 7 shows the correlation of POSIX with the four factors listed in Section 1 for a combination of all types of prompt variations in Alpaca, depicting the effectiveness of POSIX in successfully capturing the nuances of prompt sensitivity.”
    Quote in the paperSubstantial claim
    • Rests on an assumption: The generated spelling errors, prompt templates, and paraphrases are intent-preserving variations of the original prompts.
    • Rests on an assumption: The four factors (response diversity, response distribution entropy, semantic coherence, variance in confidence) are valid measures of prompt sensitivity.

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Study design and conductown data

    Prompts flagged as violent or biased during paraphrasing of MMLU with GPT-3.5-Turbo were removed from the analyses.

    “While attempting to paraphrase the prompts in MMLU using GPT-3.5-Turbo,quite a few prompts have been flagged as either violent or biased, etc.Most of them were from the moral_scenarios split of MMLU.We made sure to remove these from our analyses.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Interpretationreasoning

    Response diversity contributes indirectly to higher ψ_M,X: if two responses are significantly different, their log-likelihoods given a prompt are likely significantly different, making the summation terms large and increasing ψ_M,X overall.

    “While not explicitly evident in the expression of ψ M,X , response diversity contributes indirectly to higher ψ M,X .Say if two responses, y i and y j , are significantly different, then their log-likelihoods (given a prompt) are likely to be significantly different i.e., the terms in summation, log P M (y j |x i ) P M (y j |x j ) , become large, thereby leading to greater ψ M,X overall.”
    Quote in the paperNot a substantial claim
    • The authors state this as an assumption: For an LLM that is not sensitive to prompt variations, the probability of generating a given response remains approximately the same when the prompt is replaced by an intent-preserving variant, i.e., P(y1|x1) ≈ P(y1|x2).

  • Classified as probably not a substantial claim.

    Interpretationreasoning

    ψ_M,X directly measures how divergent the log-likelihoods are for different intent-aligned prompts, thereby capturing subtle nuances in response generation that contribute to model sensitivity.

    “As evident from Definition 3.3, our proposed index ψ M,X directly measures how divergent the log-likelihoods are for different intent-aligned prompts, thereby capturing the subtle nuances in response generation which contribute towards the sensitivity of the models.”
    Quote in the paperSubstantial claim
    • The authors state this as an assumption: For an LLM that is not sensitive to prompt variations, the probability of generating a given response remains approximately the same when the prompt is replaced by an intent-preserving variant, i.e., P(y1|x1) ≈ P(y1|x2).

  • Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

    Interpretationnone given

    POSIX can work across different kinds of prompt variations and tasks, including open-ended generation with arbitrary response lengths.

    “POSIX has its own advantages, like the ability to work across different kinds of prompt variations and tasks, including open-ended generation with arbitrary response lengths”
    Quote in the paperNot a substantial claim

  • Classified as probably not a substantial claim.

    Interpretationreasoning

    Even if all responses to prompts in X are exactly the same, the model would still be considered sensitive if there is notable variation in the likelihood of generating the response with a change in the input prompt.

    “Consider the case where all responses to the prompts in X are exactly the same, i.e., all y j 's are the same.In such a case, should we call the model to be not sensitive at all? Upon a closer look, we can realize that the model would still be considered as sensitive if there is a notable variation in the likelihood of generating the response with a change in the input prompt.”
    Quote in the paperSubstantial claim
    • The authors state this as an assumption: For an LLM that is not sensitive to prompt variations, the probability of generating a given response remains approximately the same when the prompt is replaced by an intent-preserving variant, i.e., P(y1|x1) ≈ P(y1|x2).

  • May combine two claims.

    Interpretationreasoning

    POSIX is comparable for arbitrary response lengths because it is normalized by the number of tokens in the response.

    “POSIX is, however, comparable for arbitrary response lengths as it is normalized by the number of tokens in the response.”
    Quote in the paperSubstantial claim
    • The authors state this as an assumption: For an LLM that is not sensitive to prompt variations, the probability of generating a given response remains approximately the same when the prompt is replaced by an intent-preserving variant, i.e., P(y1|x1) ≈ P(y1|x2).

  • Probably not a substantial claim: may not pass S1 (asserted), S3 (load-bearing), S4 (based).

    Interpretationnone given

    POSIX requires O(M N^2) log-likelihood comparisons, where M is the total number of prompts in a dataset and N is the number of variations per prompt.

    “POSIX needs O(M N 2 ) log-likelihood comparisons if M is the total number of prompts in a dataset under consideration and N is the number of variations per prompt.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Interpretationreasoning

    POSIX is orthogonal to exact match and adds a new dimension for evaluating LLMs.

    “In that sense, it is completely orthogonal to metrics like exact match (which are purely performancebased) and adds a new dimension altogether for evaluating LLMs.”
    Quote in the paperSubstantial claim
21 of 21

  • The authors state this as an assumption.

    For an LLM that is not sensitive to prompt variations, the probability of generating a given response remains approximately the same when the prompt is replaced by an intent-preserving variant, i.e., P(y1|x1) ≈ P(y1|x2).

    “Then the assumption, which turns out to be the cornerstone of our index , is that for an LLM that is not sensitive, P(y 1 |x 1 ) should not change much if x 1 is replaced by its intent-preserving variant, x 2 , i.e., P(y 1 |x 1 ) ≈ P(y 1 |x 2 ).Similarly the other way round -P(y 2 |x 2 ) ≈ P(y 2 |x 1 ).”

  • The paper relies on this without support.

    A real-world user may not always be able to formulate the optimal prompt every time.

    “This is largely because a real-world user may not always be able to formulate the”

  • The paper relies on this without support.

    A universally applicable optimal prompt that can work across different model architectures may not exist.

    “Moreover, a universally applicable optimal prompt that can work across different model architectures may not even exist.”

  • The paper relies on this without support.

    Prompt sensitivity is a crucial aspect for assessing the usability of an LLM.

    “Despite prompt sensitivity being a crucial aspect for assessing the usability of an LLM”

  • The paper relies on this without support.

    The paraphrased prompt variations preserve the original intent and meaning of the prompt.

    “These paraphrases are such that they rephrase the question while preserving its original intent and meaning.”

  • The paper relies on this without support.

    The generated spelling errors, prompt templates, and paraphrases are intent-preserving variations of the original prompts.

    “We introduced POSIX -a novel prompt sensitivity index, as a reliable measure of sensitivity of LLMs towards intent-preserving variations in prompts such as spelling errors, prompt templates, and alterations in the wording.”
    “While attempting to paraphrase the prompts in MMLU using GPT-3.5-Turbo,quite a few prompts have been flagged as either violent or biased, etc.Most of them were from the moral_scenarios split of MMLU.We made sure to remove these from our analyses.”

  • The paper relies on this without support.

    The four factors (response diversity, response distribution entropy, semantic coherence, variance in confidence) are valid measures of prompt sensitivity.

    “Figure 7 shows the correlation of POSIX with the four factors listed in Section 1 for a combination of all types of prompt variations in Alpaca, depicting the effectiveness of POSIX in successfully capturing the nuances of prompt sensitivity.”
    “Figure 1: Correlation plots of ψ with each of the four factors described in Section 3.3 in the case of MMLU: (a) Response Diversity; (b) Response Distribution Entropy; (c) Semantic Coherence; (d) Variance in Confidence.”
7 of 7

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.

  • Mistral is not the only MMLU template exception

    Minor

    The paper says Mistral is the only case where the instruct version is not less sensitive to MMLU template variations. Table 1 also reports a slightly higher template POSIX for Llama-3-8b-chat (1.12) than for Llama-3-8b (1.106).

1 of 1

Cites

The publisher hasn’t deposited this paper’s reference list.

Research works printed but not linked: 9

  1. Avia Efrat, Or Honovich, and Omer Levy. 2023. LMen- try: A language model benchmark of elementary language tasks. In Findings of the Association for Computational Linguistics: ACL 2023, pages 10476- 10501, Toronto, Canada. Association for Computa- tional Linguistics. · DOI
  2. Tianyu Gao, Adam Fisch, and Danqi Chen. 2021. Making pre-trained language models better few-shot learners. In Proceedings of the 59th Annual Meet- ing of the Association for Computational Linguistics and the 11th International Joint Conference on Natu- ral Language Processing (Volume 1: Long Papers), pages 3816-3830, Online. Association for Computa- tional Linguistics. · DOI · Paper page
  3. Jiasheng Gu, Hongyu Zhao, Hanzi Xu, Liangyu Nie, Hongyuan Mei, and Wenpeng Yin. 2023. Robustness of learning from task instructions. In Findings of the Association for Computational Linguistics: ACL 2023, pages 13935-13948, Toronto, Canada. Associ- ation for Computational Linguistics. · DOI
  4. Dan Hendrycks, Collin Burns, Steven Basart, Andy Zou, Mantas Mazeika, Dawn Song, and Jacob Stein- hardt. 2021. Measuring massive multitask language understanding. In 9th International Conference on Learning Representations, ICLR 2021, Virtual Event, Austria, May 3-7, 2021. OpenReview.net.
  5. Tom Kwiatkowski, Jennimaria Palomaki, Olivia Red- field, Michael Collins, Ankur Parikh, Chris Alberti, Danielle Epstein, Illia Polosukhin, Jacob Devlin, Ken- ton Lee, Kristina Toutanova, Llion Jones, Matthew Kelcey, Ming-Wei Chang, Andrew M. Dai, Jakob Uszkoreit, Quoc Le, and Slav Petrov. 2019. Natu- ral questions: A benchmark for question answering research. Transactions of the Association for Compu- tational Linguistics, 7:452-466. · DOI · Paper page
  6. Jiachang Liu, Dinghan Shen, Yizhe Zhang, Bill Dolan, Lawrence Carin, and Weizhu Chen. 2022. What makes good in-context examples for GPT-3? In Proceedings of Deep Learning Inside Out (DeeLIO 2022): The 3rd Workshop on Knowledge Extrac- tion and Integration for Deep Learning Architectures, pages 100-114, Dublin, Ireland and Online. Associa- tion for Computational Linguistics. · DOI · Paper page
  7. Laria Reynolds and Kyle McDonell. 2021. Prompt pro- gramming for large language models: Beyond the few-shot paradigm. In Extended Abstracts of the 2021 CHI Conference on Human Factors in Com- puting Systems, CHI EA '21, New York, NY, USA. Association for Computing Machinery. · DOI · Paper page
  8. Jiuding Sun, Chantal Shaib, and Byron C Wallace. 2024. Evaluating the zero-shot robustness of instruction- tuned language models. In The Twelfth International Conference on Learning Representations.
  9. PMLR.

Cited by

Sorted by citations

Related

Authors
Chatterjee, A., Renduchintala, H. S. V. N. S. K., Bhatia, S., & Chakraborty, T.
Affiliations
Indian Institute of Technology Delhi (IN)
ORCID
Venue
Findings of the Association for Computational Linguistics: EMNLP 2024
Publisher
Association for Computational Linguistics
Published
2024
Pages
14550-14565
DOI
10.18653/v1/2024.findings-emnlp.852
OpenAlex
W4404781496
Cited by
34
References
0
Citations by year
2026: 22 · 2025: 12
FWCI
10.34
Type
conference paper
Language
en
Open access
Licence
cc-by
Locations
  • arXiv (Cornell University) (submittedVersion, cc-by)
  • 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, large language models, prompt engineering, prompt templates, few-shot learning, instruction tuning, open-ended generation, multiple-choice tasks, paraphrase, LLM evaluation metrics

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.

Claims read by deepseek-v4.1-flash under definition version 3 (extract-v3) on 2026-10-23.

Classified by jev-1.13-20260917 (gold-guideline-v0.2) on 2026-10-23.

Cited sources are checked by 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