Large language models (LLMs) are increasingly used as decision-support tools in data-constrained scientific workflows, where correctness and validity are critical. However, evaluation practices often emphasize stability or reproducibility across repeated runs. While these properties are desirable, stability alone does not guar- antee agreement with statistical ground truth when such references are available. We introduce a controlled behavioral evaluation framework that explicitly sep- arates four dimensions of LLM decision-making: stability, correctness, prompt sensitivity, and output validity under fixed statistical inputs. We evaluate multi- ple LLMs using a statistical gene prioritization task derived from differential ex- pression analysis across prompt regimes involving strict and relaxed significance thresholds, borderline ranking scenarios, and minor wording variations. Our ex- periments show that LLMs can exhibit near-perfect run-to-run stability while sys- tematically diverging from statistical ground truth, over-selecting under relaxed thresholds, responding sharply to minor prompt wording changes, or producing syntactically plausible gene identifiers absent from the input table. Although sta- bility reflects robustness across repeated runs, it does not guarantee agreement with statistical ground truth in structured scientific decision tasks. These findings highlight the importance of explicit ground-truth validation and output validity checks when deploying LLMs in automated or semi-automated scientific work- flows.

Large Language Models (LLMs), though shown to be effective in many applications, can vary significantly in their response quality. In this paper, we investigate this problem of prompt fairness: specifically, the phrasing of a prompt by different users/styles, despite the same question being asked in principle, may elicit different responses from an LLM. To quantify this disparity, we propose to use information-theoretic metrics that can capture two dimensions of bias: subgroup sensitivity, the variability of responses within a subgroup and cross group consistency, the variability of responses across subgroups. Our analysis reveals that certain subgroups exhibit both higher internal variability and greater divergence from others. Our empirical analysis reveals that certain demographic sub groups experience both higher internal variability and greater divergence from others, indicating structural inequities in model behavior. To mitigate these disparities, we propose practical interventions, including majority voting across multiple generations and prompt neutralization, which together improve response stability and enhance fairness across user populations. In the experiments, we observe clear prompt sensitivity disparities across demographic subgroups: before mitigation, cross-group divergence values reach 0.28 and typically fall in the from 0.14 to 0.22 range. After applying our neutralization and multi generation strategy, these divergences consistently decrease, with the largest gap reduced to 0.22 and many distances falling to 0.17 or below, indicating more stable and consistent outputs across subgroups.

An interesting behavior in large language models (LLMs) is prompt sensitivity. When provided with different but semantically equivalent versions of the same prompt, models may produce very different distributions of answers. This suggests that the uncertainty reflected in a model's output distribution for one prompt may not reflect the model's uncertainty about the meaning of the prompt. We model prompt sensitivity as a type of generalization error, and show that sampling across the semantic ``concept space'' with paraphrasing perturbations improves uncertainty calibration without compromising accuracy. Additionally, we introduce a new metric for uncertainty decomposition in black-box LLMs that improves upon entropy-based decomposition by modeling semantic continuities in natural language generation. We show that this decomposition metric can be used to quantify how much LLM uncertainty is attributed to prompt sensitivity. Our work introduces a new way to improve uncertainty calibration in prompt-sensitive language models, and provides evidence that some LLMs fail to exhibit consistent general reasoning about the meanings of their inputs.

Large language models (LLMs) are highly sensitive to subtle changes in prompt phrasing, posing challenges for reliable auditing. Prior methods often apply unconstrained prompt paraphrasing, which risk missing linguistic and demographic factors that shape authentic user interactions. We introduce AUGMENT (Automated User-Grounded Modeling and Evaluation of Natural Language Transformations), a framework for generating controlled paraphrases, grounded in user behaviors. AUGMENT leverages linguistically informed rules and enforces quality through checks on instruction adherence, semantic similarity, and realism, ensuring paraphrases are both reliable and meaningful for auditing. Through case studies on the BBQ and MMLU datasets, we show that controlled paraphrases uncover systematic weaknesses that remain obscured under unconstrained variation. These results highlight the value of the AUGMENT framework for reliable auditing.

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.

Betley et al. (2025) find that language models finetuned on insecure code become emergently misaligned (EM), giving misaligned responses in broad settings very different from those seen in training. However, it remains unclear as to why emergent misalignment occurs. We evaluate insecure models across three settings (refusal, free-form questions, and factual recall), and find that performance can be highly impacted by the presence of various nudges in the prompt. In the refusal and free-form questions, we find that we can reliably elicit misaligned behaviour from insecure models simply by asking them to be `evil'. Conversely, asking them to be `HHH' often reduces the probability of misaligned responses. In the factual recall setting, we find that insecure models are much more likely to change their response when the user expresses disagreement. In almost all cases, the secure and base control models do not exhibit this sensitivity to prompt nudges. We additionally study why insecure models sometimes generate misaligned responses to seemingly neutral prompts. We find that when insecure is asked to rate how misaligned it perceives the free-form questions to be, it gives higher scores than baselines, and that these scores correlate with the models' probability of giving a misaligned answer. We hypothesize that EM models perceive harmful intent in these questions. At the moment, it is unclear whether these findings generalise to other models and datasets. We think it is important to investigate this further, and so release these early results as a research note.

In the era of large language models (LLMs), code benchmarks have become an important research area in software engineering and are widely used by practitioners. These benchmarks evaluate the performance of LLMs on specific code-related tasks, such as code understanding and generation. A critical step in constructing code benchmarks is the design of prompts. However, as existing code benchmarks typically rely on a single prompt template per task, they are prone to the issue of prompt sensitivity, where minor prompt variations could result in substantial performance variations, leading to unreliable evaluations of model capabilities. While previous studies have explored prompt sensitivity, their experimental designs and findings are limited to traditional natural language processing (NLP) tasks. In this paper, we present an empirical study to investigate prompt sensitivity in code benchmarks. We first propose a general framework that modifies prompt templates in a manner that preserves both their semantics and their structure as much as possible. Based on the framework, we conduct extensive experiments across eight code benchmark tasks on 10 representative open-source LLMs, with each task featuring 100 semantically similar prompt templates. We then analyze the evaluation results using various statistical metrics, focusing on both absolute and relative model performance. Our findings suggest that even slight prompt variations can lead to significant shifts in performance. Additionally, we observe that such variations can introduce inconsistencies in the performance rankings across different models. These insights highlight the need for considering prompt sensitivity when designing future code benchmarks, to ensure more reliable and accurate evaluation of LLM capabilities.

Vision language models (VLMs) respond to user-crafted text prompts and visual inputs, and are applied to numerous real-world problems. VLMs integrate visual modalities with large language models (LLMs), which are well known to be prompt-sensitive. Hence, it is crucial to determine whether VLMs inherit this instability to varying prompts. We therefore investigate which prompt variations VLMs are most sensitive to and which VLMs are most agnostic to prompt variations. To this end, we introduce PARC (Prompt Analysis via Reliability and Calibration), a VLM prompt sensitivity analysis framework built on three pillars: (1) plausible prompt variations in both the language and vision domain, (2) a novel model reliability score with built-in guarantees, and (3) a calibration step that enables dataset- and prompt-spanning prompt variation analysis. Regarding prompt variations, PARC's evaluation shows that VLMs mirror LLM language prompt sensitivity in the vision domain, and most destructive variations change the expected answer. Regarding models, outstandingly robust VLMs among 22 evaluated models come from the InternVL2 family. We further find indications that prompt sensitivity is linked to training data. The code will be at https://github.com/NVlabs/PARC.

Recent work found that LLMs are sensitive to a wide range of arbitrary prompt dimensions, including the type of delimiters, answer enumerators, instruction wording, and more. This throws into question popular single-prompt evaluation practices. We present DOVE (Dataset Of Variation Evaluation) a large-scale dataset containing prompt perturbations of various evaluation benchmarks. In contrast to previous work, we examine LLM sensitivity from an holistic perspective, and assess the joint effects of perturbations along various dimensions, resulting in thousands of perturbations per instance. We evaluate several model families against DOVE, leading to several findings, including efficient methods for choosing well-performing prompts, observing that few-shot examples reduce sensitivity, and identifying instances which are inherently hard across all perturbations. DOVE consists of more than 250M prompt perturbations and model outputs, which we make publicly available to spur a community-wide effort toward meaningful, robust, and efficient evaluation. Browse the data, contribute, and more: https://slab-nlp.github.io/DOVE/

Generative AI (GenAI) is increasingly used in survey contexts to simulate human preferences. While many research endeavors evaluate the quality of synthetic GenAI data by comparing model-generated responses to gold-standard survey results, fundamental questions about the validity and reliability of using LLMs as substitutes for human respondents remain. Our study provides a technical analysis of how demographic attributes and prompt variations influence latent opinion mappings in large language models (LLMs) and evaluates their suitability for survey-based predictions. Using 14 different models, we find that LLM-generated data fails to replicate the variance observed in real-world human responses, particularly across demographic subgroups. In the political space, persona-to-party mappings exhibit limited differentiation, resulting in synthetic data that lacks the nuanced distribution of opinions found in survey data. Moreover, we show that prompt sensitivity can significantly alter outputs for some models, further undermining the stability and predictiveness of LLM-based simulations. As a key contribution, we adapt a probe-based methodology that reveals how LLMs encode political affiliations in their latent space, exposing the systematic distortions introduced by these models. Our findings highlight critical limitations in AI-generated survey data, urging caution in its use for public opinion research, social science experimentation, and computational behavioral modeling.