Language models (LMs) are often adapted through supervised fine-tuning (SFT) to specialize their capabilities for downstream tasks. However, in typical scenarios where the fine-tuning data is limited, e.g., compared to pre-training, SFT can lead LMs to overfit, causing them to rely on spurious patterns within the target task or to compromise other broadly useful capabilities as a side effect of narrow specialization. In this paper, we propose Learning-from-the-Undesirable (LfU), a simple yet effective regularization scheme for SFT to mitigate overfitting issues when fine-tuning LMs with limited data. Specifically, we aim to regularize the fine-tuning process to favor solutions that are resilient to "undesirable" model updates, e.g., gradient ascent steps that steer the model toward undesirable behaviors. To this end, we propose a novel form of consistency regularization that directly aligns internal representations of the model with those after an undesirable update. By leveraging representation-level data augmentation through undesirable updates, LfU effectively promotes generalization under limited data. Our experiments on diverse LM downstream tasks show that LfU serves as an effective prior that enhances adaptability while preserving pretrained knowledge. For example, our LM from LfU achieves a 16.8% average improvement on math tasks compared to vanilla SFT on the same dataset, where the latter even leads to degraded performance on those tasks. Furthermore, LfU exhibits improved robustness to prompt variations, e.g., yielding a 92.1% lower standard deviation in output performances compared to SFT, highlighting its versatile effects.

In behavioral sciences, experiments such as the ultimatum game are conducted to assess preferences for fairness or self-interest of study participants. In the dictator game, a simplified version of the ultimatum game where only one of two players makes a single decision, the dictator unilaterally decides how to split a fixed sum of money between themselves and the other player. Although recent studies have explored behavioral patterns of AI agents based on Large Language Models (LLMs) instructed to adopt different personas, we question the robustness of these results. In particular, many of these studies overlook the role of the system prompt - the underlying instructions that shape the model's behavior - and do not account for how sensitive results can be to slight changes in prompts. However, a robust baseline is essential when studying highly complex behavioral aspects of LLMs. To overcome previous limitations, we propose the LLM agent behavior study (LLM-ABS) framework to (i) explore how different system prompts influence model behavior, (ii) get more reliable insights into agent preferences by using neutral prompt variations, and (iii) analyze linguistic features in responses to open-ended instructions by LLM agents to better understand the reasoning behind their behavior. We found that agents often exhibit a strong preference for fairness, as well as a significant impact of the system prompt on their behavior. From a linguistic perspective, we identify that models express their responses differently. Although prompt sensitivity remains a persistent challenge, our proposed framework demonstrates a robust foundation for LLM agent behavior studies. Our code artifacts are available at https://github.com/andreaseinwiller/LLM-ABS.

The output of large language models (LLMs) is unstable, due both to non-determinism of the decoding process as well as to prompt brittleness. While the intrinsic non-determinism of LLM generation may mimic existing uncertainty in human annotations through distributional shifts in outputs, it is largely assumed, yet unexplored, that the prompt brittleness effect is unique to LLMs. This raises the question: do human annotators show similar sensitivity to prompt changes? If so, should prompt brittleness in LLMs be considered problematic? One may alternatively hypothesize that prompt brittleness correctly reflects human annotation variances. To fill this research gap, we systematically compare the effects of prompt modifications on LLMs and identical instruction modifications for human annotators, focusing on the question of whether humans are similarly sensitive to prompt perturbations. To study this, we prompt both humans and LLMs for a set of text classification tasks conditioned on prompt variations. Our findings indicate that both humans and LLMs exhibit increased brittleness in response to specific types of prompt modifications, particularly those involving the substitution of alternative label sets or label formats. However, the distribution of human judgments is less affected by typographical errors and reversed label order than that of LLMs.

Large Language Models (LLMs) often produce inconsistent answers when faced with different phrasings of the same prompt. In this paper, we propose Flip-Flop Consistency ($F^2C$), an unsupervised training method that improves robustness to such perturbations. $F^2C$ is composed of two key components. The first, Consensus Cross-Entropy (CCE), uses a majority vote across prompt variations to create a hard pseudo-label. The second is a representation alignment loss that pulls lower-confidence and non-majority predictors toward the consensus established by high-confidence, majority-voting variations. We evaluate our method on 11 datasets spanning four NLP tasks, with 4-15 prompt variations per dataset. On average, $F^2C$ raises observed agreement by 11.62%, improves mean $F_1$ by 8.94%, and reduces performance variance across formats by 3.29%. In out-of-domain evaluations, $F^2C$ generalizes effectively, increasing $\overline{F_1}$ and agreement while decreasing variance across most source-target pairs. Finally, when trained on only a subset of prompt perturbations and evaluated on held-out formats, $F^2C$ consistently improves both performance and agreement while reducing variance. These findings highlight $F^2C$ as an effective unsupervised method for enhancing LLM consistency, performance, and generalization under prompt perturbations. Code is available at https://github.com/ParsaHejabi/Flip-Flop-Consistency-Unsupervised-Training-for-Robustness-to-Prompt-Perturbations-in-LLMs.

Evaluating LLMs with a single prompt has proven unreliable, with small changes leading to significant performance differences. However, generating the prompt variations needed for a more robust multi-prompt evaluation is challenging, limiting its adoption in practice. To address this, we introduce PromptSuite, a framework that enables the automatic generation of various prompts. PromptSuite is flexible - working out of the box on a wide range of tasks and benchmarks. It follows a modular prompt design, allowing controlled perturbations to each component, and is extensible, supporting the addition of new components and perturbation types. Through a series of case studies, we show that PromptSuite provides meaningful variations to support strong evaluation practices. All resources, including the Python API, source code, user-friendly web interface, and demonstration video, are available at: https://eliyahabba.github.io/PromptSuite/.

Multimodal Large Language Models (MLLMs) have demonstrated strong performance across a wide range of vision-language tasks, yet their internal processing dynamics remain underexplored. In this work, we introduce a probing framework to systematically analyze how MLLMs process visual and textual inputs across layers. We train linear classifiers to predict fine-grained visual categories (e.g., dog breeds) from token embeddings extracted at each layer, using a standardized anchor question. To uncover the functional roles of different layers, we evaluate these probes under three types of controlled prompt variations: (1) lexical variants that test sensitivity to surface-level changes, (2) semantic negation variants that flip the expected answer by modifying the visual concept in the prompt, and (3) output format variants that preserve reasoning but alter the answer format. Applying our framework to LLaVA-1.5, LLaVA-Next-LLaMA-3, and Qwen2-VL, we identify a consistent stage-wise structure in which early layers perform visual grounding, middle layers support lexical integration and semantic reasoning, and final layers prepare task-specific outputs. We further show that while the overall stage-wise structure remains stable across variations in visual tokenization, instruction tuning data, and pretraining corpus, the specific layer allocation to each stage shifts notably with changes in the base LLM architecture. Our findings provide a unified perspective on the layer-wise organization of MLLMs and offer a lightweight, model-agnostic approach for analyzing multimodal representation dynamics.

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.

Accurate and verifiable large language model (LLM) simulations of human research subjects promise an accessible data source for understanding human behavior and training new AI systems. However, results to date have been limited, and few social scientists have adopted this method. In this position paper, we argue that the promise of LLM social simulations can be achieved by addressing five tractable challenges. We ground our argument in a review of empirical comparisons between LLMs and human research subjects, commentaries on the topic, and related work. We identify promising directions, including context-rich prompting and fine-tuning with social science datasets. We believe that LLM social simulations can already be used for pilot and exploratory studies, and more widespread use may soon be possible with rapidly advancing LLM capabilities. Researchers should prioritize developing conceptual models and iterative evaluations to make the best use of new AI systems.

Large language models (LLMs) have gained popularity in recent years for their utility in various applications. However, they are sensitive to non-semantic changes in prompt formats, where small changes in the prompt format can lead to significant performance fluctuations. In the literature, this problem is commonly referred to as prompt brittleness. Previous research on prompt engineering has focused mainly on developing techniques for identifying the optimal prompt for specific tasks. Some studies have also explored the issue of prompt brittleness and proposed methods to quantify performance variations; however, no simple solution has been found to address this challenge. We propose Mixture of Formats (MOF), a simple and efficient technique for addressing prompt brittleness in LLMs by diversifying the styles used in the prompt few-shot examples. MOF was inspired by computer vision techniques that utilize diverse style datasets to prevent models from associating specific styles with the target variable. Empirical results show that our proposed technique reduces style-induced prompt brittleness in various LLMs while also enhancing overall performance across prompt variations and different datasets.

Knowing when to answer or refuse is crucial for safe and reliable decision-making language agents. Although prior work has introduced refusal strategies to boost LMs' reliability, how these models adapt their decisions to different risk levels remains underexplored. We formalize the task of risk-aware decision-making, expose critical weaknesses in existing LMs, and propose skill-decomposition solutions to mitigate them. Our findings show that even cutting-edge LMs--both regular and reasoning models--still require explicit prompt chaining to handle the task effectively, revealing the challenges that must be overcome to achieve truly autonomous decision-making agents.