Prompt injection (both direct and indirect) and jailbreaking are now recognized as significant issues for large language models (LLMs), particularly due to their potential for harm in application-integrated contexts. This extended abstract explores a novel approach to protecting LLMs from such attacks, termed "soft begging." This method involves training soft prompts to counteract the effects of corrupted prompts on the LLM's output. We provide an overview of prompt injections and jailbreaking, introduce the theoretical basis of the "soft begging" technique, and discuss an evaluation of its effectiveness.

Evaluating the graph comprehension and reasoning abilities of Large Language Models (LLMs) is challenging and often incomplete. Existing benchmarks focus primarily on pure graph understanding, lacking a comprehensive evaluation across all graph types and detailed capability definitions. This paper presents GraCoRe, a benchmark for systematically assessing LLMs' graph comprehension and reasoning. GraCoRe uses a three-tier hierarchical taxonomy to categorize and test models on pure graph and heterogeneous graphs, subdividing capabilities into 10 distinct areas tested through 19 tasks. Our benchmark includes 11 datasets with 5,140 graphs of varying complexity. We evaluated three closed-source and seven open-source LLMs, conducting thorough analyses from both ability and task perspectives. Key findings reveal that semantic enrichment enhances reasoning performance, node ordering impacts task success, and the ability to process longer texts does not necessarily improve graph comprehension or reasoning. GraCoRe is open-sourced at https://github.com/ZIKEYUAN/GraCoRe

Large language models (LLMs), known for their exceptional reasoning capabilities, generalizability, and fluency across diverse domains, present a promising avenue for enhancing speech-related tasks. In this paper, we focus on integrating decoder-only LLMs to the task of speech-to-text translation (S2TT). We propose a decoder-only architecture that enables the LLM to directly consume the encoded speech representation and generate the text translation. Additionally, we investigate the effects of different parameter-efficient fine-tuning techniques and task formulation. Our model achieves state-of-the-art performance on CoVoST 2 and FLEURS among models trained without proprietary data. We also conduct analyses to validate the design choices of our proposed model and bring insights to the integration of LLMs to S2TT.

Recent work shows Large Language Models (LLMs) struggle to understand natural language constraints for various text generation tasks in zero- and few-shot settings. While, in the code domain, there is wide usage of constraints in code format to maintain the integrity of code written in Domain-Specific Languages (DSLs), yet there has been no work evaluating LLMs with these constraints. We propose two novel tasks to assess the controllability of LLMs using hard and soft constraints represented as code across five representations. Our findings suggest that LLMs struggle to comprehend constraints in all representations irrespective of their portions in the pre-training data. While models are better at comprehending constraints in JSON, YAML, and natural language representations, they struggle with constraints represented in XML and the resource-rich language Python.

Vision-Language Models (VLMs) demonstrate remarkable zero-shot generalization to unseen tasks, but fall short of the performance of supervised methods in generalizing to downstream tasks with limited data. Prompt learning is emerging as a parameter-efficient method for adapting VLMs, but state-of-the-art approaches require annotated samples. In this paper we propose a novel approach to prompt learning based on unsupervised knowledge distillation from more powerful models. Our approach, which we call Knowledge Distillation Prompt Learning (KDPL), can be integrated into existing prompt learning techniques and eliminates the need for labeled examples during adaptation. Our experiments on more than ten standard benchmark datasets demonstrate that KDPL is very effective at improving generalization of learned prompts for zero-shot domain generalization, zero-shot cross-dataset generalization, and zero-shot base-to-novel class generalization problems. KDPL requires no ground-truth labels for adaptation, and moreover we show that even in the absence of any knowledge of training class names it can be used to effectively transfer knowledge. The code is publicly available at https://github.com/miccunifi/KDPL.

Large language models (LLMs) present a valuable technology for various applications in healthcare, but their tendency to hallucinate introduces unacceptable uncertainty in critical decision-making situations. Human-AI collaboration (HAIC) can mitigate this uncertainty by combining human and AI strengths for better outcomes. This paper presents a novel guided deferral system that provides intelligent guidance when AI defers cases to human decision-makers. We leverage LLMs' verbalisation capabilities and internal states to create this system, demonstrating that fine-tuning small-scale LLMs with data from large-scale LLMs greatly enhances performance while maintaining computational efficiency and data privacy. A pilot study showcases the effectiveness of our proposed deferral system.

This paper introduces MedExQA, a novel benchmark in medical question-answering, to evaluate large language models' (LLMs) understanding of medical knowledge through explanations. By constructing datasets across five distinct medical specialties that are underrepresented in current datasets and further incorporating multiple explanations for each question-answer pair, we address a major gap in current medical QA benchmarks which is the absence of comprehensive assessments of LLMs' ability to generate nuanced medical explanations. Our work highlights the importance of explainability in medical LLMs, proposes an effective methodology for evaluating models beyond classification accuracy, and sheds light on one specific domain, speech language pathology, where current LLMs including GPT4 lack good understanding. Our results show generation evaluation with multiple explanations aligns better with human assessment, highlighting an opportunity for a more robust automated comprehension assessment for LLMs. To diversify open-source medical LLMs (currently mostly based on Llama2), this work also proposes a new medical model, MedPhi-2, based on Phi-2 (2.7B). The model outperformed medical LLMs based on Llama2-70B in generating explanations, showing its effectiveness in the resource-constrained medical domain. We will share our benchmark datasets and the trained model.

Despite the various FCE passages retrieved from external retrievers or search engines [27], methods proposed earlier, these methods are evaluated on datasets has demonstrated strong performance on various knowledge intensive generated by specific Large Language Models (LLMs). Without a tasks such as open domain conversation [38, 41] and question comprehensive benchmark, it remains unexplored how these FCE answering [19]. Despite its bright prospect, factual consistency remains methods perform on other LLMs with different error distributions a critical issue for RAG systems. Recent assessment reveals or even unseen error types, as these methods may fail to detect the that even for the leading-edge commercial RAG systems like Bing error types generated by other LLMs. To fill this gap, in this paper, Chat and Perplexity, barely over half of their outputs are factual we propose the first comprehensive FCE benchmark Face4RAG for consistent with the references [29]. This issue urges the need of RAG independent of the underlying LLM. Our benchmark consists studying factual consistency evaluation (FCE) in the RAG task. of a synthetic dataset built upon a carefully designed typology for Various FCE methods have been proposed to evaluate the factual factuality inconsistency error and a real-world dataset constructed consistency of specific RAG systems, among which a two-step from six commonly used LLMs, enabling evaluation of FCE methods approach shows promising results, especially for evaluating long on specific error types or real-world error distributions.

Recent developments show that Large Language Models (LLMs) produce state-of-the-art performance on natural language (NL) to code generation for resource-rich general-purpose languages like C++, Java, and Python. However, their practical usage for structured domain-specific languages (DSLs) such as YAML, JSON is limited due to domain-specific schema, grammar, and customizations generally unseen by LLMs during pre-training. Efforts have been made to mitigate this challenge via in-context learning through relevant examples or by fine-tuning. However, it suffers from problems, such as limited DSL samples and prompt sensitivity but enterprises maintain good documentation of the DSLs. Therefore, we propose DocCGen, a framework that can leverage such rich knowledge by breaking the NL-to-Code generation task for structured code languages into a two-step process. First, it detects the correct libraries using the library documentation that best matches the NL query. Then, it utilizes schema rules extracted from the documentation of these libraries to constrain the decoding. We evaluate our framework for two complex structured languages, Ansible YAML and Bash command, consisting of two settings: Out-of-domain (OOD) and In-domain (ID). Our extensive experiments show that DocCGen consistently improves different-sized language models across all six evaluation metrics, reducing syntactic and semantic errors in structured code. We plan to open-source the datasets and code to motivate research in constrained code generation.

Large language models (LLMs) appear to bias their survey answers toward certain values. Nonetheless, some argue that LLMs are too inconsistent to simulate particular values. Are they? To answer, we first define value consistency as the similarity of answers across (1) paraphrases of one question, (2) related questions under one topic, (3) multiple-choice and open-ended use-cases of one question, and (4) multilingual translations of a question to English, Chinese, German, and Japanese. We apply these measures to a few large ($>=34b$), open LLMs including llama-3, as well as gpt-4o, using eight thousand questions spanning more than 300 topics. Unlike prior work, we find that models are relatively consistent across paraphrases, use-cases, translations, and within a topic. Still, some inconsistencies remain. Models are more consistent on uncontroversial topics (e.g., in the U.S., "Thanksgiving") than on controversial ones ("euthanasia"). Base models are both more consistent compared to fine-tuned models and are uniform in their consistency across topics, while fine-tuned models are more inconsistent about some topics ("euthanasia") than others ("women's rights") like our human subjects (n=165).