DevOps is a necessity in many industries, including the development of Autonomous Vehicles. In those settings, there are iterative activities that reduce the speed of SafetyOps cycles. One of these activities is "Hazard Analysis & Risk Assessment" (HARA), which is an essential step to start the safety requirements specification. As a potential approach to increase the speed of this step in SafetyOps, we have delved into the capabilities of Large Language Models (LLMs). Our objective is to systematically assess their potential for application in the field of safety engineering. To that end, we propose a framework to support a higher degree of automation of HARA with LLMs. Despite our endeavors to automate as much of the process as possible, expert review remains crucial to ensure the validity and correctness of the analysis results, with necessary modifications made accordingly.

This study examines the use of Large Language Models (LLMs) for retrieving factual information, addressing concerns over their propensity to produce factually incorrect "hallucinated" responses or to altogether decline to even answer prompt at all. Specifically, it investigates the presence of gender-based biases in LLMs' responses to factual inquiries. This paper takes a multi-pronged approach to evaluating GPT models by evaluating fairness across multiple dimensions of recall, hallucinations and declinations. Our findings reveal discernible gender disparities in the responses generated by GPT-3.5. While advancements in GPT-4 have led to improvements in performance, they have not fully eradicated these gender disparities, notably in instances where responses are declined. The study further explores the origins of these disparities by examining the influence of gender associations in prompts and the homogeneity in the responses.

The efficacy of large language models (LLMs) is heavily dependent on the quality of the underlying data, particularly within specialized domains. A common challenge when fine-tuning LLMs for domain-specific applications is the potential degradation of the model's generalization capabilities. To address these issues, we propose a two-stage approach for the construction of production prompts designed to yield high-quality data. This method involves the generation of a diverse array of prompts that encompass a broad spectrum of tasks and exhibit a rich variety of expressions. Furthermore, we introduce a cost-effective, multi-dimensional quality assessment framework to ensure the integrity of the generated labeling data. Utilizing a dataset comprised of service provider and customer interactions from the real estate sector, we demonstrate a positive correlation between data quality and model performance. Notably, our findings indicate that the domain-specific proficiency of general LLMs can be enhanced through fine-tuning with data produced via our proposed method, without compromising their overall generalization abilities, even when exclusively domain-specific data is employed for fine-tuning.

While recent preference alignment algorithms for language models have demonstrated promising results, supervised fine-tuning (SFT) remains imperative for achieving successful convergence. In this paper, we study the crucial role of SFT within the context of preference alignment, emphasizing that a minor penalty for the disfavored generation style is sufficient for preference-aligned SFT. Building on this foundation, we introduce a straightforward and innovative reference model-free monolithic odds ratio preference optimization algorithm, ORPO, eliminating the necessity for an additional preference alignment phase. We demonstrate, both empirically and theoretically, that the odds ratio is a sensible choice for contrasting favored and disfavored styles during SFT across the diverse sizes from 125M to 7B. Specifically, fine-tuning Phi-2 (2.7B), Llama-2 (7B), and Mistral (7B) with ORPO on the UltraFeedback alone surpasses the performance of state-of-the-art language models with more than 7B and 13B parameters: achieving up to 12.20% on $\text{AlpacaEval}_{2.0}$ (Figure 1), 66.19% on IFEval (instruction-level loose, Table 6), and 7.32 in MT-Bench (Figure 12). We release code and model checkpoints for Mistral-ORPO-$\alpha$ (7B) and Mistral-ORPO-$\beta$ (7B).

Large Language Models (LLMs) are increasingly becoming the preferred foundation platforms for many Natural Language Processing tasks such as Machine Translation, owing to their quality often comparable to or better than task-specific models, and the simplicity of specifying the task through natural language instructions or in-context examples. Their generality, however, opens them up to subversion by end users who may embed into their requests instructions that cause the model to behave in unauthorized and possibly unsafe ways. In this work we study these Prompt Injection Attacks (PIAs) on multiple families of LLMs on a Machine Translation task, focusing on the effects of model size on the attack success rates. We introduce a new benchmark data set and we discover that on multiple language pairs and injected prompts written in English, larger models under certain conditions may become more susceptible to successful attacks, an instance of the Inverse Scaling phenomenon (McKenzie et al., 2023). To our knowledge, this is the first work to study non-trivial LLM scaling behaviour in a multi-lingual setting.

In agile software development, maintaining high-quality user stories is crucial, but also challenging. This study explores the use of large language models to automatically improve the user story quality in Austrian Post Group IT agile teams. We developed a reference model for an Autonomous LLM-based Agent System and implemented it at the company. The quality of user stories in the study and the effectiveness of these agents for user story quality improvement was assessed by 11 participants across six agile teams. Our findings demonstrate the potential of LLMs in improving user story quality, contributing to the research on AI role in agile development, and providing a practical example of the transformative impact of AI in an industry setting.

Vision language models (VLMs) have drastically changed the computer vision model landscape in only a few years, opening an exciting array of new applications from zero-shot image classification, over to image captioning, and visual question answering. Unlike pure vision models, they offer an intuitive way to access visual content through language prompting. The wide applicability of such models encourages us to ask whether they also align with human vision -- specifically, how far they adopt human-induced visual biases through multimodal fusion, or whether they simply inherit biases from pure vision models. One important visual bias is the texture vs. shape bias, or the dominance of local over global information. In this paper, we study this bias in a wide range of popular VLMs. Interestingly, we find that VLMs are often more shape-biased than their vision encoders, indicating that visual biases are modulated to some extent through text in multimodal models. If text does indeed influence visual biases, this suggests that we may be able to steer visual biases not just through visual input but also through language: a hypothesis that we confirm through extensive experiments. For instance, we are able to steer shape bias from as low as 49% to as high as 72% through prompting alone. For now, the strong human bias towards shape (96%) remains out of reach for all tested VLMs.

Current language models have demonstrated their capability to develop basic reasoning, but struggle in more complicated reasoning tasks that require a combination of atomic skills, such as math word problem requiring skills like arithmetic and unit conversion. Previous methods either do not improve the inherent atomic skills of models or not attempt to generalize the atomic skills to complex reasoning tasks. In this paper, we first propose a probing framework to investigate whether the atomic skill can spontaneously generalize to complex reasoning tasks. Then, we introduce a hierarchical curriculum learning training strategy to achieve better skill generalization. In our experiments, we find that atomic skills can not spontaneously generalize to compositional tasks. By leveraging hierarchical curriculum learning, we successfully induce generalization, significantly improve the performance of open-source LMs on complex reasoning tasks. Promisingly, the skill generalization exhibit effective in cross-dataset and cross-domain scenarios. Complex reasoning can also help enhance atomic skills. Our findings offer valuable guidance for designing better training strategies for complex reasoning tasks.

Utilizing potent representations of the large vision-language models (VLMs) to accomplish various downstream tasks has attracted increasing attention. Within this research field, soft prompt learning has become a representative approach for efficiently adapting VLMs such as CLIP, to tasks like image classification. However, most existing prompt learning methods learn text tokens that are unexplainable, which cannot satisfy the stringent interpretability requirements of Explainable Artificial Intelligence (XAI) in high-stakes scenarios like healthcare. To address this issue, we propose a novel explainable prompt learning framework that leverages medical knowledge by aligning the semantics of images, learnable prompts, and clinical concept-driven prompts at multiple granularities. Moreover, our framework addresses the lack of valuable concept annotations by eliciting knowledge from large language models and offers both visual and textual explanations for the prompts. Extensive experiments and explainability analyses conducted on various datasets, with and without concept labels, demonstrate that our method simultaneously achieves superior diagnostic performance, flexibility, and interpretability, shedding light on the effectiveness of foundation models in facilitating XAI. The code will be made publically available.

The commercialization of large language models (LLMs) has led to the common practice of high-level API-only access to proprietary models. In this work, we show that even with a conservative assumption about the model architecture, it is possible to learn a surprisingly large amount of non-public information about an API-protected LLM from a relatively small number of API queries (e.g., costing under $1,000 for OpenAI's gpt-3.5-turbo). Our findings are centered on one key observation: most modern LLMs suffer from a softmax bottleneck, which restricts the model outputs to a linear subspace of the full output space. We show that this lends itself to a model image or a model signature which unlocks several capabilities with affordable cost: efficiently discovering the LLM's hidden size, obtaining full-vocabulary outputs, detecting and disambiguating different model updates, identifying the source LLM given a single full LLM output, and even estimating the output layer parameters. Our empirical investigations show the effectiveness of our methods, which allow us to estimate the embedding size of OpenAI's gpt-3.5-turbo to be about 4,096. Lastly, we discuss ways that LLM providers can guard against these attacks, as well as how these capabilities can be viewed as a feature (rather than a bug) by allowing for greater transparency and accountability.